International Journal of Computer Science and Security Volume (4) Issue (2)

Editor in Chief Dr. Haralambos Mouratidis

International Journal of Computer
Science and Security (IJCSS)
Book: 2010 Volume 4, Issue 2
Publishing Date: 30-05-2010
Proceedings
ISSN (Online): 1985-1553

This work is subjected to copyright. All rights are reserved whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting,
re-use of illusions, recitation, broadcasting, reproduction on microfilms or in any
other way, and storage in data banks. Duplication of this publication of parts
thereof is permitted only under the provision of the copyright law 1965, in its
current version, and permission of use must always be obtained from CSC
Publishers. Violations are liable to prosecution under the copyright law.

IJCSS Journal is a part of CSC Publishers
http://www.cscjournals.org

© IJCSS Journal
Published in Malaysia

Typesetting: Camera-ready by author, data conversation by CSC Publishing
Services – CSC Journals, Malaysia

CSC Publishers

Table of Contents

Volume 4, Issue 2, May 2010.

Pages
149 - 159 Parallel Computing in Chemical Reaction Metaphor with Tuple
Space
Hong Lin, Jeremy Kemp, Wilfredo Molin

160 - 175 A novel Data Mining algorithm for semantic web based data cloud
N.C.Mahanti, kanhaiya lal

176 - 182 Document Topic Generation in Text Mining by Using Cluster
Analysis with EROCK
Rizwan Ahmad, Aasia Khanum

183 - 198 Heuristics Based Genetic Algorithm for Scheduling Static Tasks in
Homogeneous Parallel System
Kamaljit Kaur, Amit Chhabra, Gurvinder Singh

199 - 207 Knowledge Discovery from Students’ Result Repository:
Association Rule Mining Approach
Olanrewaju Jelili Oyelade, Oladipupo, Olufunke Oyejoke

208 - 225 Multilevel Access Control in a MANET for a Defense Messaging
System Using Elliptic Curve Cryptography
J. Nafeesa Begum, K.Kumar, V.Sumathy

226 - 236 FPGA Prototype of Robust Image Watermarking For JPEG 2000
With Dual Detection
Pankaj U.Lande, Sanjay N. Talbar, G.N. Shinde

237 - 255 Comparing The Proof By Knowledge Authentication Techniques
Stamati Gkarafli , Anastasios A. Economides

256 - 264 Development of Information Agent Reranking By Using Weights
Measurement
Aliaa A.Youssif, Ashraf A. Darwish, Ahmed Roshdy

International Journal of Computer Science and Security (IJCSS), Volume (4), Issue (2)

Hong Lin, Jeremy Kemp & Wilfredo Molina

Parallel Computing in Chemical Reaction Metaphor with Tuple
Space

Hong Lin linh@uhd.edu
Department of Computer and Mathematical Sciences
University of Houston-Downtown
1 Main Street, Houston, Texas 77002, USA

Jeremy Kemp

Wilfredo Molina

Abstract

Methodologies have been developed to allow parallel programming in a higher
level. These include the Chemical Reaction Models, Linda, and Unity. We
present the Chemical Reaction Models and its implementation in IBM Tuple
Space. Sample programs have been developed to demonstrate this
methodology.

Keywords: Parallel Programming, Very High Level Languages, the Chemical Reaction Model, IBM Tuple
Space.

1. Higher-Level Parallel Computing - Implicit Parallelism
Higher level parallel programming models express parallelism in an implicit way. Instead of
imposing programmers to create multiple tasks that can run concurrently and handle their
communications and synchronizations explicitly, these models allow programs to be written
without assumptions of artificial sequenciality. The programs are naturally parallel. Examples of
such kind of models include the Chemical Reaction Models (CRMs) [1, 2], Linda [3], and Unity [4,
5]. These models are created to address higher level programming issues such as formal
program specification, program synthesis, program derivation and verification, and software
architecture. Efficient implementation of these models has limited success and therefore
obscures its direct applications in software design [6, 7]. Despite this limitation, efforts have been
made in both academic and industrial settings to avail these models in real-world programming.
For example, Unity has been used in industrial software design and found successful; execution
efficiency of Linda has been affirmed by experiments and it is implemented by IBM Tuple Space.
Recent discussions of these models in multi-agent system design have also been found in
literature [8]. In the following discussion, we focus on the Chemical Reaction Models and its
applications.

The Chemical Reaction Models describe computation as “chemical reactions”. Data (the
“solution”) are represented as a multiset. A set of “reaction” rules is given to combine elements in

International Journal of Computer Science and Security (IJCSS), Volume (4): Issue (2) 149


the multiset and produce new elements. Reactions take place until the solution becomes inert,
namely there are no more elements can be combined. The results of computation are
represented as the inert multiset. Gamma is a kernel language in which programs are described
in terms of multiset transformations. In Gamma programming paradigm, programmers can
concentrate on the logic of problem solving based on an abstract machine and are free from
considering any particular execution environment. It has seeded follow-up elaborations, such as
Chemical Abstract Machine (Cham) [9], higher-order Gamma [10, 11], and Structured Gamma
[12]. While the original Gamma language is a first-order language, higher order extensions have
been proposed to enhance the expressiveness of the language. These include higher-order
Gamma, hmm-calculus, and others. The recent formalisms, γ-Calculi, of Gamma languages
combine reaction rules and the multisets of data and treat reactions as first-class citizens [13-15].
Among γ-Calculi, γ0-Calculus is a minimal basis for the chemical paradigm; γc-Calculus extends
γ0-Calculus by adding a condition term into γ-abstractions; and γn-Calculus extends γ0-Calculus
by allowing abstractions to atomically capture multiple elements. Finally, γcn-Calculus combines
both γc-Calculus and γn-Calculus. For notational simplicity, we use γ-Calculus to mean γcn-
Calculus from this point on.

The purpose of the presented study is to investigate a method for implementing γ-Calculus using
IBM Tuple Space. TSpace supports network computation in client/server style. The target of this
effort is to enable higher order programming in a parallel computing platform, such as computer
clusters, and allow for refinement of the executable programs using transformation techniques.

The paper will be organized as follows. In Section 2, we give a brief introduction to γ-Calculus. In
Section 3, we discuss the method for implementing γ-Calculus in IBM Tuple space. Program
examples are presented in Section 4. We conclude in Section 5.

2. γ-Calculus
The basic term of a Gamma program is molecules (or γ-expressions), which can be simple data
or programs (γ-abstractions). The execution of the Gamma program can be seen as the evolution
of a solution of molecules, which react until the solution becomes inert. Molecules are recursively
defined as constants, γ-abstractions, multisets or solution of molecules. The following is their
syntax:

M ::= 0 | 1 | … | ‘a’ | ‘b’ | … ; constants
| γP[C].M ; γ-abstraction
| M 1, M 2 ; multiset
| <M> ; solution

The multiset constructor “,” is associative and commutative (AC rule). Solutions encapsulate
molecules. Molecules can move within solutions but not across solutions. γ-abstractions are
elements of multisets, just like other elements. They can be applied to other elements of the
same solution if a match to pattern P is found and condition C evaluates to true and therefore
facilitate the chemical reaction. The pattern has the following syntax:

P ::= x | P, P | <P>

where x is a variable. In addition, we allow for the use of tuples (written x 1:… : xn) and names of
types. For example, γ-abstraction

γ(x: Int, y: Int)[x ≥ y].x

can be interpreted as: replace x, y by x if x ≥ y, which is equivalent to finding the maximum of
two integers.



The semantics of γ-Calculus is defined as the following:

(γp[c].m1), m2 = фm1 if match(p/ m2) = ф and фc ; γ-conversion
m1, m2 = m 2, m 1 ; commutativity
m1, (m2, m3) = (m1, m2), m3 ; associativity
E1 = E2 => E[E1] = E[E2] ; chemical law

The γ-conversion describes the reaction mechanism. When the pattern p matches m2, a
substitution ф is yielded. If the condition фc holds, the reactive molecules γp[c].m1 and m2 are
consumed and a new molecule фm1 is produced. match(p/m) returns the substitution
corresponding to the unification of variables if the matching succeeds, otherwise it returns fail.

Chemical law formalizes the locality of reactions. E[E1] denotes the molecule obtained by
replacing holes in the context E[ ] (denoted by [ ]) by the molecule E1. A molecule is said to be
inert if no reaction can be made within:

Inert(m)
(m ≡ m’[(γp[c].m1), m2] => match(p/m2) = fail)

A solution is inert if all molecules within are inert and normal forms of chemical reactions are inert
γ-expression. Elements inside a solution can be matched only if the solution is inert. Therefore, a
pattern cannot match an active solution. This ensures that solutions cannot be decomposed
before they reach their normal form and therefore permits the sequentialization of reactions. The
following inference rule governs the evolution of γ-expressions:

E1  E 2 E  C[ E1] E '  C[ E 2]
E  E'
This high level language can be implemented in Java using IBM's TSpaces server and Java
package. The method is detailed in the following section.

3. IBM Tuple Space
IBM Tuple Space was originally invented as an implementation of Linda computational model.
While version 3.0 is only available after obtaining a site license, version 2.12 is freely available.
Installing TSpaces is as easy as unpackaging the TSpaces package on the networked file system
(NFS), adding it's directory to the users classpath, and starting the server in a GNU Screen
session.

3.1 Data Structures and Methods
A TSpaces program uses more of a client/server model, with each node only communicating with
a 'host' or any machine with a known name, which on a cluster would usually be the head node.
And although TSpaces is flexible enough to assign global names and ranks to each node, micro-
managing the communications, this would defeat the purpose of having the abstraction layer
TSpaces offers you. Data in a TSpaces program is shared through Tuples, a tuple is a data
structure that wraps all other data structures in a tuplespace, this can include data primitives, as
well as standard java classes and user defined classes. Every tuple is in a TuplesSpace, and
every TupleSpace has a host (actually a fully qualified host name).

The TSpaces methods used to obtain tuples are: read(), waitToRead(), take(), waitToTake(). The
read() and take() methods vary in that the read() method leaves the tuple returned to the program
in the tuplespace, and the take() method removes it from the tuplespace. The waitTo versions
wait until a tuple appears in a tuplespace, and then takes it; these are good for synchronization,
and can be set to time out in order to prevent the program from hanging indefinitely. These



methods take Tuples as arguments, and return the first tuple from the tuplespace that matches
either the data types you specified, or specific values. There is also the scan() method, which
works like read() except it returns all values that match the tuple specified. There are other
Tspace methods that allow you more control over your program; countN() returns all the matching
tuples and delete() removes all matching tuples. There are many other advanced features that
allow you to manipulate TSpaces with a great deal of precision.

Another bonus of using a platform that runs on Java is being able to run these programs on any
OS, and even inside a web browser using an applet. Such applets are ideal for being able to
monitor or visualize a reaction that is running.

3.2 Synchronization
The most difficult task in writing a CRM/gamma program in Tspaces is synchronization.
Synchronization is needed to determine when to terminate the program, or when the molecules
stop reacting.

One program acts as one molecule, and different programs can read from the same tuplespace,
but in order to test that the reaction of a molecule is exhausted, a program must run through the
entire data space. So, reactions that involve more then one molecule type should be run in a
cycle, the first program reacting all of the molecules from the original tuplespace, and writing the
resulting tuples, as well as all of the unused tuples to the next tuplespace. The next program acts
on this tuplespace until it is stable with regard to the molecule it represents, leaving the tuples in a
third tuplespace. This continues until all of the programs that represent molecules have had a
chance to run, and then it starts over. Termination occurs when the programs cycle through all of
the programs without any changes to the tuples.

Molecules that reduce tuplespaces to a single solution are the simplest. They require only one
tuplespace and no synchronization. The max number finder and the tuple adder are examples of
this, they simply react until there is only one element left, then they are done.

After some further thought (but not further programming) a relation of numbers of tuplespaces to
input and output elements of the gamma function can be noticed. In order to use the style of
synchronization used in the sorting program, a program with X input molecules and Y output
molecules requires X tuplespaces for the input and Y tuplespaces for the output. In order to
detect stability, it will have to empty the input tuples 1 time with no changes and empty the output
tuplespace 1 time.

The sorting program is an example of this; the method it uses involves alternating between
comparing two tuples of the form (even index n, dataValue), (n +1, dataValue) and (odd index m,
dataValue), (m+1, dataValue), and incrementing a counter tuple whenever a change occurs. This
allows a comparison after the last tuples are removed from the tuplespace and into another
tuplespace to determine if any changes have been made. If two consecutive changeless runs
occur, then every data element is in order, and the program terminates.

There is a complication with programs where number of inputs is not equal to the number of
outputs. The indexing must be handles specially or removed entirely; with numbers removed,
there will be n + 1 elements missing, and with numbers added, there will be elements that need to
be added somewhere, these can usually be appended to the end?

If there are more inputs then outputs, then the tuplespace will eventually be reduced to the
number of outputs for the one molecule. These types of programs can use the termination style of
the max element and tuple adder programs; simply running until there are Y elements left, and
then stopping. The only synchronization required is when emptying a tuplespace(or set of
tuplespaces), and to prevent deadlock when (number of elements) < (number of processors) *
(number of inputs), but this can be handled by detecting null reads and random timeouts on the
waitToRead() calls.



Although Tuple Space was initially implemented to support Linda computation model, its functions
well suite in the operational semantics of the chemical reactions models. We propose
implementing γ-Calculus on Tuple Space. In the following, we demonstrate a few sample
programs in γ-Calculus and their Tuple Space implementation.

4. Examples

4.1 Max Number Finder
Given a set of values of an ordered type M, this program returns the maximum number of the set.
The following γ-abstraction compares two randomly selected values. If the first value is greater
than or equal to the second, it removes the second value from the set:

select = γ(a: M, b: M)[a ≥ b]. a: M, select

No global control is imposed on the way multiset elements are selected to ignite the reaction. If
select is placed in an initial set M0 of values, it will compare two values and erase the smaller at a
time till the maximum is left. So the maximum number program can then be written as:

Max M0 = <select, M0>

If the multiset M0 is represented as a tuple space, this program can be converted into one that
finds and displays the greatest tuple inside a tuple space. It works with each node taking two
tuples from the tuple space, comparing them, and placing the greatest one back to the tuple
space. This process repeats itself until the termination condition is met, that is, when there is only
one tuple left in the tuple space. When a node picks tuples up, if both tuples happen to be the
same size, it simply places one of them back in the tuplespace while discarding the other one. If a
node happens to take only one tuple because another node already picked the last remaining
ones in the tuple space, this puts it back and repeats the process. This ensures that by the next
check, a node will be able to take two tuples and perform the remaining operations to find the
greatest tuple. If a node sees no tuples in the tuple space, this displays a message and
terminates. If a node sees only one tuple in the tuple space, it assumes the greatest tuple was
already found, displays a message and terminates.

Check appendix A for an example of the code implemented in TSpaces as well as its flowchart.

4.2 Tuple Adder
Given a set of values of numerical type M, we write a program to summarize all the values. The
following γ-abstraction adds two randomly selected values and put the sum back into the multiset:

add = γ(a: M, b: M)[true]. a+b: M, select

The tuple adder program can then be written as:

Sum M0 = <add, M0>

If M0 is represented as a tuple space, the corresponding TSpace program will add all of the tuples
in a tuple space and displays their total sum. It works with each node taking two random tuples
from the tuple space, adding them up, and placing a new tuple with their total sum back in the
tuplespace (the tuples themselves are discarded). This process repeats itself until there is only
one tuple left in the tuplespace, which is the total sum. If there are no tuples in the tuplespace
before execution, the nodes display a message and terminate.



Check appendix B for the flowchart of the code implemented in TSpaces. Code is omitted
because of the page limit.

4.3 Sorter
If a list is represented by multiset M = {(a, i) | a is value and i an index and i’s are consecutive},
the following recursive γ-abstraction replaces any ill-ordered pairs by two other pairs:

sigma = γ((a, i): M, (b, j): M)[ i < j  a > b]. (b, i): M, (a, j): M, sigma

It specifies that any two selected pairs (a, i) and (b, j) that satisfy the condition, i < j a > b are
replaced by two other pairs (b, i) and (a, j), and a copy of itself. If sigma is placed in an initial set
M0 of pairs, it will replace ill-ordered pairs until the entire list is sorted. So a sorting program can
be defined as:

Sort M0 = <sigma, M0>

In a tuple space, a similar process will happen. The program will sort all of the tuples in a tuple
space in ascending order. Each tuple has an index and a value in the following format: (index,
value). When two tuples, (i, x) and (j, y) from said tuple space S are taken by a node, it first
checks whether x > y & i < j. If this happens to be true, then the following swap is performed: (i,
y), (j, x) they are put back in the tuple space, and the tuples are in order. This process repeats
itself until no more swaps can be performed, that is, when all of the tuples in a tuple space are
arranged in ascending order.

As mentioned above, multiple tuplespaces are required to synchronize this 'single pool'
abstraction, in this case four tuplespaces were used. There is a primary pool, where the data is
initially stored and an alternate pool where the data is written as it is being processed. Each of
these pools is broken in to an even and an odd pool

Check appendix C for the flowchart of the code implemented in TSpaces, the primary feature of
this programming model, is that it can utilize up to n/2 processing nodes, where n is the number
of data items being sorted.

We have tested the above TSpace programs on a PC cluster and observed the computation in
multiple nodes, and how the increase of nodes divides the number of comparisons per node, and
increases speed; all of this thanks to the abstractions and portability offered by TSpaces.

We want to point out that when converting a γ-Calculus program into a TSpace program, details
must be added to make a working program. However, through the above examples, we can see
the conversion is straightforward in sense of the computation model. Therefore, it is technically
feasible to design an automatic conversion system that can parse γ-Calculus and convert the
program into a TSpace program and this is the next goal of our ongoing project.

5. Conclusion
The Chemical Reaction Models are higher level programming models that address parallelism
implicitly and allows programmers to focus on the logic of problem solving instead of deal with
operational details in parallel computing. IBM Tuple Space supports client/server computation
based on Linda model that uses a similar concept for data structures. We discuss a method for
implementing a higher order chemical reaction model, γ-Calculus, in IBM Tuple Space. We
present the rules for converting γ-Calculus constructs into TSpace codes and discuss the critical
techniques such as synchronizations. Our work shows that the conversion is practical.
Experiments are also conducted on a computer cluster.



6. Acknowledgements
This research is partially supported by NSF grant “Acquisition of a Computational Cluster Grid for
Research and Education in Science and Mathematics” (#0619312).

7. References
1. Banatre, J.-P. and Le Metayer, D. “The Gamma model and its discipline of programming”.
Science of Computer Programming, 15, 55-77, 1990.
2. Banatre, J.-P. and Le Metayer, D. “Programming by multiset transformation”. CACM, 36(1),
98-111, 1993.
3. Carriero, N. and Gelernter, D. “Linda in context”. CACM, 32(4), 444-458, 1989.
4. K.M. Chandy and J. Misra. “Parallel Program Design: A Foundation”, Addison-Wesley (1988)
5. Misra, J. “A foundation of parallel programming”. In M. Broy (ed.), Constructive Methods in
Computing Science. NATO ASI Series, Vol. F55, 397-443, 1989.
6. C. Creveuil. “Implementation of Gamma on the Connection Machine”. In Proc. Workshop on
Research Directions in High-Level Parallel Programming Languages, Mont-Saint Michel,
1991, Springer-Verlag, LNCS 574, 219-230, 1991.
7. Gladitz, K. and Kuchen, H. “Shared memory implementation of the Gamma-operation”.
Journal of Symbolic Computation 21, 577-591, 1996.
8. Cabri, et al. “Mobile-Agent Coordination Models for Internet Applications”. Computer, 2000
February, http://dlib.computer.org/co/books/co2000/pdf/r2082.pdf. 2000.
9. Berry, G. and Boudol, G. “The Chemical Abstract Machine”. Theoretical Computer Science,
96, 217-248, 1992.
10. Le Metayer, D. “Higher-order multiset processing”. DIMACS Series in Discrete Mathematics
and Theoretical Computer Science, 18, 179-200, 1994.
11. Cohen, D. and Muylaert-Filho, J. “Introducing a calculus for higher-order multiset
programming”. In Coordination Languages and Models, LNCS, 1061, 124-141, 1996.
12. Fradet, P. and Le Metayer, D. “Structured Gamma”. Science of Computer Programming,
31(2-3), 263-289, 1998.
13. J.-P. Banâtre, P. Fradet and Y. Radenac. “Chemical specification of autonomic systems”. In
Proc. of the 13th International Conference on Intelligent and Adaptive Systems and Software
Engineering (IASSE'04), July 2004.
14. J.-P. Banâtre, P. Fradet and Y. Radenac. “Principles of chemical programming”. In S.
Abdennadher and C. Ringeissen (eds.): Proc. of the 5th International Workshop on Rule-
Based Programming (RULE'04), 124, ENTCS, 133-147, 2005.
15. J.-P. Banâtre, P. Fradet and Y. Radenac. “Higher-order Chemical Programming Style”. In
Proceedings of Unconventional Programming Paradigms, Springer-Verlag, LNCS, 3566, 84-
98, 2005.



Appendix A – Maximum Number Finder

Flowchart

Code

//Wilfredo Molina – July 10, 2009

import com.ibm.tspaces.*;

public class gammaMax3
{
public static void main(String[] args)
{
try
{
String host = "grid.uhd.edu";
TupleSpace ts = new TupleSpace("gammaSort", host);
Tuple template = new Tuple(new Field(Integer.class), new Field(Double.class));

Tuple t1;
Tuple t2;

if ((Integer)ts.countN(template) < 1)
System.out.println("TupleSpace Empty Here");
else
{
while ((Integer)ts.countN(template) > 1)
{
t1 = (Tuple)ts.take(template);
t2 = (Tuple)ts.take(template);

if (t1 == null || t2 == null)
{
if (t1 != null)



ts.write(t1);
if (t2 != null)
ts.write(t2);
}
else
{
if ((Double)t1.getField(1).getValue() >
(Double)t2.getField(1).getValue())
ts.write(t1);
else
ts.write(t2);
}
}

if ((Integer)ts.countN(template) == 1)
{
t1 = (Tuple)ts.read(template);
System.out.println("Max Found: " +
(Double)t1.getField(1).getValue());
}
}
}
catch (TupleSpaceException tse)
{
System.out.println("It's Broke, Wil.");
}
}
}



Appendix B – Tuple Adder



Appendix C – Sorter

Start

Do counters Do counters
Make counter in TS
exist in eTS? exist in OTS?

Wait to read counter
End Term flag?
in TS

Move to oOTS

Does counter W ait for counter to
Data in eTS? Data in OTS Switch Tuplespaces
exist? be written

Wait for all elements
Data in OTS? Move eTS to eOTS to be checked in Write counter

Get Tuple [i] from Does change
eTS Set NCCR to 0 Switch Tuplespaces
counter = 0 ?

Get Tuple [i+1]
from oTS
Does NCCR =
Increment NCCR Set term flag
2?

Remove checkout

Increment checkout

Write checkout

Is Tuple [i] > Switch Tuple [i+1] Wait to remove
Tuple [i+1] and Tuple [i] counter

Increment

Remove checkout Write counter

Decrement checkout

Buy SmartDraw !- purchased copies print this
document without a watermark .
Write checkout
Visit www .smartdraw.com or call 1-800-768-3729.


Kanhaiya Lal & N.C.Mahanti

A Novel Data Mining Algorithm for Semantic Web Based Data
Cloud

Kanhaiya Lal klal@bitmesra.ac.in
Sr. Lecturer/Dept. of Computer Sc. & Engg.,
Birla Institute of Technology, Patna Campus,
Patna, 800014, India

N.C.Mahanti ncmahanti@rediffmail.com
Professor & Head/Department of Applied Mathematics
Birla Institute of Technology, Mesra
Ranchi, 835215, India

Abstract

By a cloud, we mean an infrastructure that provides resources and/or services
over the Internet. A storage cloud provides storage services, while a compute
cloud provides compute services. We describe the design of the Sector storage
cloud and how it provides the storage services required by the Sphere compute
cloud [14]. Different efforts have been made to address the problem of data
mining in the cloud framework. In this paper we propose an algorithm to mine the
data from the cloud using sector/sphere framework and association rules. We
also describe the programming paradigm supported by the Sphere compute
cloud and Association rules. Sector and Sphere are discussed for analyzing large
data sets using computer clusters connected with wide area high performance
networks

Keywords: Cloud, Web, Apriori, Sphere, Association, Sector, confidence, support.

1. INTRODUCTION
Data mining is a treatment process to extract useful and interesting knowledge from large amount
of data. The knowledge modes data mining discovered have a variety of different types. The
common patterns are: association mode, classification model, class model, sequence pattern and
so on.
Mining association rules is one of the most important aspects in data mining. Association rules
are dependency rules which predict occurrence of an item based on occurrences of other items. It
is simple but effective and can help the commercial decision making like the storage layout,
appending sale and etc. We usually use distributed system as a solution to mining association
rules when mass data is being collected and warehoused. With the development of web and
distributed techniques, we begin to store databases in distributed systems. Thus researches on
the algorithm of mining association rules in distributed system are becoming more important and
have a broad application foreground. Distributed algorithm has characters of high adaptability,
high flexibility, low wearing performance and easy to be connected etc. [15].

The Semantic Web is based on a vision of Tim Berners-Lee, the inventor of the WWW. The great
success of the current WWW leads to a new challenge: A huge amount of data is interpretable by



humans only; machine support is limited. Berners-Lee suggests enriching the Web by machine-
process able information which supports the user in his tasks. For instance, today’s search
engines are already quite powerful, but still too often return excessively large or inadequate lists
of hits. Machine process able information can point the search engine to the relevant pages and
can thus improve both precision and recall. For instance, today it is almost impossible to retrieve
information with a keyword search when the information is spread Fig. 1. The layers of the
Semantic Web over several pages. Consider, e.g., the query for Web Mining experts in a
company intranet, where the only explicit information stored are the relationships between
people and the courses they attended on one hand, and between courses and the topics they
cover on the other hand. In that case, the use of a rule stating that people who attended a course
which was about a certain topic have knowledge about that topic might improve the results. The
process of building the Semantic Web is currently an area of high activity. Its structure has to be
defined, and this structure then has to be filled with life. In order to make this task feasible, one
should start with the simpler tasks first. The following steps show the direction where the
Semantic Web is heading:
1. Providing a common syntax for machine understandable statements.
2. Establishing common vocabularies.
3. Agreeing on a logical language.
4. Using the language for exchanging proofs.
Berners-Lee suggested a layer structure for the Semantic Web. This structure reflects the steps
listed above. It follows the understanding that each step alone will already provide added value,
so that the Semantic Web can be realized in an incremental fashion.[17]

Figure 1. The layers of the Semantic Web.

The paper reveals a distributed high performance data mining system called Sector/Sphere that
is based on an entirely different paradigm. Sector is designed to provide long term persistent
storage to large datasets that are managed as distributed indexed files. In this paper, it has been
described the design of Sector/Sphere. We also describe a data mining application developed
using Sector/Sphere that searches for emergent behavior in distributed network data. Different
segments of the file are scattered throughout the distributed storage managed by Sector. Sector
generally replicates the data to ensure its longevity, to decrease the latency when retrieving it.
and to provide opportunities for parallelism. Sector is designed to take advantage of wide area
high performance networks when available [14]. The data is persistently stored and processed in
place whenever possible. In this model, the data waits for the task or query. The storage clouds
provided by Amazon’s S3 [1], the Google File System [2], and the open source Hadoop
Distributed File System (HDFS) [3] support this model. With the Sector/Sphere software from
Source Forge, Terasort and Terasplit benchmarks, and the Angle datasets from the Large Data
Archive, the algorithm may be implemented.



2. BACKGROUND & RELATED WORK
Cloud means, an infrastructure that provides resources and/or services over the Internet. A
storage cloud provides storage services (block or file based services); a data cloud provides data
management services (record-based, column-based or object-based services); and a compute
cloud provides computational services. Often these are layered (compute services over data
services over storage service) to create a stack of cloud services that serves as a computing
platform for developing cloud-based applications [14].
Examples include Google’s Google File System (GFS), BigTable and MapReduce infrastructure
[4]; Amazon’s S3 storage cloud, SimpleDB data cloud, and EC2 compute cloud [5]; and the open
source Hadoop system [3]. In this section, we describe some related work in high performance
and distributed data mining. For a recent survey of high performance and distributed data mining
systems, see [6].
By and large, data mining systems that have been developed for clusters, distributed clusters and
grids have assumed that the processors are the scarce resource, and hence shared. When
processors become available, the data is moved to the processors, the computation is started,
and results are computed and returned [7]. In practice with this approach, for many computations,
a good portion of the time is spent transporting the data.
Key Characteristics
Agility: Agility improves with users' ability to rapidly and inexpensively re-provision technological
infrastructure resources.
Cost: Cost is claimed to be greatly reduced and capital expenditure is converted to operational
expenditure. This ostensibly lowers barriers to entry, as infrastructure is typically provided by a
third-party and does not need to be purchased for one-time or infrequent intensive computing
tasks. Pricing on a utility computing basis is fine-grained with usage-based options and fewer IT
skills ,are required for implementation (in-house).
Device and location independence enable users to access systems using a web browser
regardless of their location or what device they are using (e.g., PC, mobile). As infrastructure is
off-site (typically provided by a third-party) and accessed via the Internet, users can connect from
anywhere.

Multi-tenancy enables sharing of resources and costs across a large pool of users. One of the
most compelling reasons for vendors/ISVs to utilize multi-tenancy is for the inherent data
aggregation benefits. Instead of collecting data from multiple data sources, with potentially
different database schemas, all data for all customers is stored in a single database schema.
Thus, running queries across customers, mining data, and looking for trends is much simpler.
This reason is probably overhyped as one of the core multi-tenancy requirements is the need to
prevent Service Provider access to customer (tenant) information.

Centralization of infrastructure in locations with lower costs (such as real estate, electricity,
etc.)
Peak-load capacity increases highest possible load-levels.
Utilization and efficiency improvements for systems that are often only 10–20% utilized.
Reliability improves through the use of multiple redundant sites, which makes cloud computing
suitable for business continuity and disaster recovery. Nonetheless, many major cloud computing
services have suffered outages, and IT and business managers can at times do little when they
are affected.
Scalability via dynamic ("on-demand") provisioning of resources on a fine-grained, self-service
basis near real-time, without users having to engineer for peak loads. Performance is monitored



and consistent and loosely-coupled architectures are constructed using web services as the
system interface.
Security could improve due to centralization of data, increased security-focused resources, etc.,
but concerns can persist about loss of control over certain sensitive data, and the lack of security
for stored kernels. Security is often as good as or better than under traditional systems, in part
because providers are able to devote resources to solving security issues that many customers
cannot afford. Providers typically log accesses, but accessing the audit logs themselves can be
difficult or impossible. Furthermore, the complexity of security is greatly increased when data is
distributed over a wider area and / or number of devices.
Sustainability comes through improved resource utilization, more efficient systems, and carbon
neutrality. Nonetheless, computers and associated infrastructure are major consumers of energy.
Maintenance cloud computing applications are easier to maintain, since they don't have to be
installed on each user's computer. They are easier to support and to improve since the changes
reach the clients instantly.

FIGURE 2: Cloud Computing Model

Cloud Computing Models
Cloud computing is a highly touted recent phenomenon. As noted, there is little hope of obtaining
consensus or a standard definition regarding exactly what constitutes a ‘‘cloud” (and the term
‘‘grid” has been similarly overloaded). For example, emphasizes quality of service contracts for a
cloud, contrasts social issues with technical infrastructure, while others focus on price or on the
nature of the resources provided (e.g., storage, processors, platforms, or application services).
Some writers emphasize what the cloud provides to its consumers, e.g., services on demand.
Others emphasize what is underneath—a warehouse full of servers. The following features,
especially the first three, are commonly associated with clouds. A consumer can be an individual
lab, a consortium participant, or a consortium. _ Resource outsourcing: Instead of a consumer
providing their own hardware, the cloud vendor assumes responsibility for hardware acquisition
and maintenance. _ Utility computing: The consumer requests additional resources as needed,
and similarly releases these resources when they are not needed. Different clouds offer different
sorts of resources, e.g., processing, storage, management software, or application services .
Large numbers of machines: Clouds are typically constructed using large numbers of inexpensive
machines.
As a result, the cloud vendor can more easily add capacity and can more rapidly replace
machines that fail, compared with having machines in multiple laboratories. Generally speaking



these machines are as homogeneous as possible both in terms of configuration and location. _
Automated resource management: This feature encompasses a variety of configuration tasks
typically handled by a system administrator. For example, many clouds offer the option of
automated backup and archival. The cloud may move data or computation to improve
responsiveness. Some clouds monitor their offerings for malicious activity. _ Virtualization:
Hardware resources in clouds are usually virtual; they are shared by multiple users to improve
efficiency. That is, several lightly-utilized logical resources can be supported by the same physical
resource. _ Parallel computing: Map/Reduce and Hadoop are frameworks for expressing and
executing easily-parallelizable computations, which may use hundreds or thousands of
processors in a cloud.[16]
To enable a holistic enterprise-modeling of software assets and facilitate the generalization of
services across an organization and even beyond its boundaries, the Service-oriented modeling
framework (SOMF) offers virtualization capabilities.

FIGURE 3: SOMF Cloud Computing Model

This ability to abstract services in spite of their location, interoperability challenges, or contribution
to an architecture model fosters an elastic Cloud Computing Environment (CCE) that is nimble
enough to adapt to changes and vital to business or technological imperatives.
Moreover, the Service-oriented modeling framework (SOMF) as an enterprise modeling language
generalizes services . Thus, the notion of “Everything-as-a-Service” encompasses the cloud
computing distributed entity model (as illustrated on the far right): infrastructure-as-a-Service
(IaaS), platform-as-a-service (PaaS), and software-as-a-service (SaaS). These can be modeled
by SOMF as it conceives a service as an abstracted organizational asset and not necessarily as
a Web service.
Modeling a cloud computing not only requires a language that must be able to abstract services
and an environment that is typically virtual, but also hide the implementation from consumers.
SOMF offers these abstraction capabilities by elevating the abstraction level of an organizational
asset to enable higher cloud computing reusability rates.
Types by visibility:



Public cloud
Public cloud or external cloud describes cloud computing in the traditional mainstream sense,
whereby resources are dynamically provisioned on a fine-grained, self-service basis over the
Internet, via web applications/web services, from an off-site third-party provider who shares
resources and bills on a fine-grained utility computing basis
Hybrid cloud
A hybrid cloud environment consisting of multiple internal and/or external providers will be typical
for most enterprises. A hybrid cloud can describe configuration combining a local device, such as
a Plug computer with cloud services. It can also describe configurations combining virtual and
physical, collocated assets—for example, a mostly virtualized environment that requires
physical servers, routers, or other hardware such as a network appliance acting as a firewall or
spam filter
Private cloud
Private cloud and internal cloud are neologisms that some vendors have recently used to
describe offerings that emulate cloud computing on private networks. These products claim to
deliver some benefits of cloud computing without the pitfalls, capitalising on data security,
corporate governance, and reliability concerns. They have been criticized on the basis that users
"still have to buy, build, and manage them" and as such do not benefit from lower up-front capital
costs and less hands-on management, essentially lacking the economic model that makes cloud
computing such an intriguing concept.
While an analyst predicted in 2008 that private cloud networks would be the future of corporate
IT, there is some uncertainty whether they are a reality even within the same firm. Analysts also
claim that within five years a "huge percentage" of small and medium enterprises will get most of
their computing resources from external cloud computing providers as they "will not have
economies of scale to make it worth staying in the IT business" or be able to afford private
clouds. Analysts have reported on Platform's view that private clouds are a stepping stone to
external clouds, particularly for the financial services, and that future data centres will look like
internal clouds.
The term has also been used in the logical rather than physical sense, for example in reference
to platform as service offerings, though such offerings including Microsoft's Azure Services
Platform are not available for on-premises deployment.

FIGURE 4: Cloud Computing Types



3. ASSOCIATION RULES.
Definition 1 confidence
Set up I={i1,i2,im}for items of collection, for item in ij(1≤j≤m), (1≤j≤m) for lasting item, D={T1,TN} it is
a trade collection, Ti⊆I (1≤i≤N) here T is the trade. Rule r → q is probability that concentrates
on including in the trade.
The association rule here is an implication of the form r → q where X is the conjunction of
conditions, and Y is the type of classification. The rule r →q has to satisfy specified minimum
support and minimum confidence measure .The support of Rule r → q is the measure of
frequency both r and q in D S(r) = |r|/|D|
The confidence measure of Rule r → q is for the premise that includes r in the bargain descend,
in the meantime includes q C(r → q) =S(rq) /S(r)
Definition 2 Weighting support
Designated ones project to collect I = {i1, i2, im}, each project ij is composed with the value wj of
right(0≤ j ≤1, 1≤j ≤m). If the rule is r → q, the weighting support is

And, the K is the size of the Set rq of the project. When the right value wj is the same as ij, we are
calculating the weighting including rule to have the same support.

Association rule mining is the current hot. In association rule mining algorithms, the most
algorithms are based on Apriori algorithm to calculate, and in the mining process they can
produce amount of option set, which reduce the efficiency of the association rule mining; at the
same time the association rule mining will obtain amount of redundant rules, which will reduce the
validity of the association rule mining; and also the user interaction performance of the
association rule mining is relatively poor. On the basis of in-depth study on the existing data
mining algorithms, according to the disadvantages of the association rule mining algorithms in the
relational databases, a new data mining algorithm based on association rule is presented.
Existing mining algorithm of association rules can be broadly divided into search algorithms,
hierarchical algorithms, data sets partitioning algorithm, and so on[11] .

1) Search algorithms
Search algorithm is to deal with all the term sets contained in the affairs which were read into the
data set, so it needs to calculate the support of all term sets in the data aggregate D. Search
algorithm can find all the frequent term sets only through one scan on data sets, an affair
contained n projects will generate 2n −1 term sets, and when the terms the data set D contained
is very large, the quantity of the option sets should be calculated and stored is often very large.
Therefore, such algorithms can only be applied in the association rule mining with relatively
concentrative data.

2) Hierarchy algorithm



The hierarchy algorithm with Apriori algorithm to be representation is to find frequent term sets
since childhood and until now in the terms contained. Apriori algorithm will find all the frequent k
term sets in the first k scan on the data sets, the option sets in the first k +1 scan will be
generated by all frequent k term sets through connecting computing. The number Apriori
algorithm need to scan data sets is equal to the term numbers of the maximum frequent term
sets. The hierarchical algorithm with Apriori algorithm to be representation can generate a
relatively small option set, and the number of scanning database is decided by the term numbers
of the maximum frequent term sets.

3) Partitioning algorithm
Data set partitioning algorithm includes partition algorithm, DIC algorithm will divide the entire
data set into data blocks which can be stored in memory to handle, in order to save the I/O
spending of visiting rendering. Partition algorithm only requires two times of scan on the entire
data set. DIC algorithm can identify all the frequent term sets through the two times of scan when
appropriately dividing data blocks. The number of the option term set of the data set dividing the
algorithm generally larger than it of Apriori algorithm, increasing the data distortion can reduce the
number of the option term set. Data set partitioning algorithm is the basis of the various parallel
association rule mining algorithm and distributed association rule mining algorithm.

4) Apriori Algorithm

Apriori algorithm is an effective algorithm to mine Boolean association rule frequent term sets.
Apriori algorithm uses the strategy of breadth-first search, that is, layer-by-layer search iterative
method, first of all, find out the frequent term set with length of 1 which is recorded as L1 , L1 is
used to find the aggregate L2 of frequent 2-term sets, L2 is used to find the aggregate L3 of
frequent 3-term sets, and so the cycle continues, until no new frequent k - term sets can be
found. Finding each Lk needs a database scan. Finding all the frequent term sets is the core of
association rule mining algorithm and it has the maximum calculating workload. Afterward,
according to the minimum confidence threshold the effective association rules can be constructed
from the frequent term sets[10].

5) Sphere
The Sphere Compute Cloud is designed to be used with the Sector Storage Cloud. Sphere is
designed so that certain specialized, but commonly occurring, distributed computing operations
can be done very simply. Specifically, if a user defines a function p on a distributed data set a
managed by Sector, then invoking the command
sphere.run(a, p);

applies the user defined function p to each data record in the dataset a. In other words, if the
dataset a contains 100, 000, 000 records a[i], then the Sphere command above replaces all the
code required to read and write the array a[i] from disk, as well as the loop:

for (int i = 0, i < 100000000; ++i)
p(a[i]);



The Sphere programming model is a simple example of what is commonly called a stream
programming model. Although this model has been used for some time, it has recently received
renewed attention due to its use by the general purpose GPU (Graphics Processing Units)
community. Large data sets processed by Sphere are assumed to be broken up into several files.
For example, the Sloan Digital Sky Survey dataset [12] is divided up into 64 separate files, each
about 15.6 GB in size. The files are named sdss1.dat, . . ., sdss64.dat. Assume that the user has
a written a function called find-BrownDwarf that given a record in the SDSS dataset, extracts
candidate Brown Dwarfs. Then to find brown dwarfs in the Sloan dataset, one uses the following
Sphere code:

Stream sdss;
sdss.init(...); //init with 64 sdss files
Process* myproc = Sector::createJob();
myproc->run(sdss, "findBrownDwarf");
myproc->read(result);

With this code, Sphere uses Sector to access the required SDSS files, uses an index to extract
the relevant records, and for each record invokes the user defined function find- BrownDwarf.
Parallelism is achieved in two ways. First, the individual files can be processed in parallel.
Second, Sector is typically configured to create replicas of files for archival purposes. These
replicas can also be processed in parallel. An important advantage provided by a system such as
Sphere is that often data can be processed in place, without moving it. In contrast, a grid system
generally transfers the data to the processes prior to processing [7].

FIGURE 5: Sphere operators process Sphere streams over distributed Sphere Processing Elements (SPE)
[14].



6) Sector

Sector storage cloud described as designed for wide area, high performance 10 Gb/s networks
and employs specialized protocols, such as UDT to utilize the available bandwidth on these
networks. Sector provides services that rely in part on the local native file systems.

The characteristics of sector are
1. Sector is designed to support a community of users, not all of whom may have write access to
the Sector infrastructure.
2. Sector provides long term archival storage and access for large distributed datasets.
3. Sector is designed to utilize the bandwidth available on wide area high performance networks.
4. Sector supports a variety of different routing and network protocols. [14].

Sector has separate layers for routing and transport and interfaces with these layers through well
defined APIs. In this way, it is relatively straightforward to use other routing or network protocols.
In addition, UDT is designed in such a way that a variety of different network protocols can be
used simply by linking in one of several different libraries. As an example, Sector is used to
archive and to distribute the Sloan Digital Sky Survey (SDSS) to astronomers around the world.
Using Sector, the SDSS BESTDR5 catalog, which is about 1.3TB when compressed, can be
transported at approximately 8.1 Gb/s over a 10 Gb/s wide area network with only 6 commodity
servers [14].

Sector assumes that large datasets are divided into multiple files, say file01.dat, file02.dat, etc. It
also assumes that each file is organized into records. In order to randomly access a record in the
data set, each data file in Sector has a companion index file, with a post-fix of “.idx”. Continuing
the example above, there would be index files file01.dat.idx, file02.dat.idx, etc. The data file and
index file are always co-located on the same node. Whenever Sector replicates the data file, the
index file is also replicated. The index contains the start and end positions (i.e., the offset and
size) of each record in the data file. For those data files without an index, Sphere can only
process them at the file level, and the user must write a function that parses the file and extracts
the data
A Sector client accesses data using Sector as follows:
1. The Sector client connects to a known Sector server S, and requests the locations of an entity
managed by Sector using the entity’s name.
2. The Sector Server S runs a look-up inside the server network using the services from the
routing layer and returns one or more locations to the client. In general, an entity managed by
Sector is replicated several times within the Sector network. The routing layer can use information
involving network bandwidth and latency to determine which replica location should be provided
to the client.

3. The client requests a data connection to one or more servers on the returned locations using a
specialized Sector library designed to provide efficient message passing between geographically
distributed nodes. The Sector library used for messaging uses a specialized protocol developed
for Sector called the Group Messaging Protocol.



4. All further requests and responses are performed using a specialized library for high
performance network transport called UDT [8]. UDT is used over the data connection established
by the message passing library.

FIGURE 6: Sector consist of several layered services

Sector is designed to support a variety of different routing and networking protocols. The version
used for the experiments described below are designed to support large distributed datasets, with
loose management provided by geographically distributed clusters connected by a high
performance wide area network. With this configuration, a peer-to-peer routing protocol (the
Chord protocol described in [9]) is used so that nodes can be easily added and removed from the
system. The next version of Sector will support specialized routing protocols designed for wide
area clouds with uniform bandwidth and approximately equal RTT between clusters, as well as
non-uniform clouds in which bandwidth and RTT may vary widely between different clusters of
the cloud. Data transport within Sector is done using specialized network protocols. In particular,
data channels within Sector use high performance network transport protocols, such as UDT [13].
UDT is a rate-based application layer network transport protocol that supports large data flows
over wide area high performance networks. UDT is fair to several large data flows in the sense
that it shares bandwidth equally between them. UDT is also friendly to TCP flows in the sense
that it backs off when congestion occurs, enabling any TCP flows sharing the network to use the
bandwidth they require. Message passing with Sector is done using a specialized network
transport protocol developed for this purpose called the Group Messaging Protocol or GMP. [14].

4. METHODOLOGY
Steps for data mining using sector sphere and association rules
1. Select the minimum support threshold(Ts) and minimum confidence threshold(Tc), minimum
data size(Sizemin) and maximum data size (Sizemax).
2. We now input the data stream to the sphere processing elements. The stream is divided
into data segments. The number of data segments per SPE is calculated on the basis of
number of SPE and the entire stream size. Data segments from the same file are not
processed at the same time until other SPE become idle.
3. The SPE accepts a new data segment from the client, which contains the file name, offset,
number of rows to be processed, and additional parameters.
4. The SPE reads the data segment and its record index from local disk or from a remote disk
managed by sector.
5. For each data segment find out the frequent term set with length of 1 which is recorded as
L1 , L1 is used to find the aggregate L2 of frequent 2-term sets, L2 is used to find the



aggregate L3 of frequent 3-term sets, and so the cycle continues, until no new frequent k -
term sets can be found.
6. We generate strong association rules on the basis of the found frequent term sets i.e. we
generate those association rules whose support and confidence respectively greater than
or equal to the pre-given support threshold (Ts) and confidence threshold (Tc).
7. For each data segment (single data record, group of data records, or entire data file), the
Sphere operator processes the data segment using the association rules and writes the
result to a temporary buffer. In addition, the SPE periodically sends acknowledgments and
feedback to the client about the progress of the processing.
8. When the data segment is completely processed, the SPE sends an acknowledgment to
the client and writes the results to the appropriate destinations, as specified in the output
stream. If there are no more data segments to be processed, the client closes the
connection to the SPE, and the SPE is released.

Pseudocode for data mining using sector sphere & association rules

procedure data_mining_cloud()
{
Set minimum data segment size Dmin
Set maximum data segment size Dmax
Set minimum support threshold Ts
Set maximum confidence threshold Tc
Size of stream=S
Number of SPE=N
Nd= S/N;
for(i=0;i<N;i++)
{
Ndi’ =0;
}

spawn(SPE);
for each SPE while there are data segments to be processed
{
start a SPE connection
If(Ndi’ > Ndi)
return;
Accept new data segment in Data[D max)
metadata[8]=substr(Data, D max -8, D max );
findFrequentTermSets(Data, D max -8);
sendAcknowledgement(Ndi’);
Ndi’ = Ndi’ +1;
release SPE connection
return;
}

}

procedure findFrequentTermSets(Data, D max -8) // method 1//
{
For(k=2;Lk-1≠| |); k++)
{
Ck=apriori_gen(Lk-1 );
for each element t€ Data



{
Ct=subset(Ck,t);
for each candidate c€ Ct
c.count++;
end for;

}
Lk={c € Ck | c.count>Ts}
}
return L=U k L k;
}

procedure apriori_gen(L k-1 ){
for each itemset l1 €L k-1
for each itemset l2 €L k-1
if(l1[1]=l2[1])^ (l1[2]=l2[2]) ])^ (l1[3]=l2[3]) ])^ (l1[4]=l2[4])^...... ])^ (l1[k-2]=l2[k-2]) ])^ (l1[k-1]=l2[k-
1]) then{
c=l1 join l2;
if has_infrequent_subset(c,Lk-1) then
delete c;
else add c to C k;
}
end for
end for
return C k ;
}

procedure has_infrequent_subset(c,Lk-1)
{
for each k-1 subset of s of c
If S not belongs to Lk-1 then
return TRUE;
return FALSE;
}

//Method 2//
For( i = 1; i < n ; i++ )
{
Read in partition pi (pi belongs to p);
i
L =gen_i_length_itemsets(pi);
f(x);//shifting partitions
f(y);//shifting contents in pipeline stages to next stage
}

Procedure for gen_i_length_itemsets(pi)
p
L1 ={ large 1- itemsets along with their tidlists }
For( k = 1 ; k ≤ m ; k++ )
{

For all itemsets l1 Lk-1p do begin

p
For all itemsets l2 Lk-1 do begin



If ( l1[1]=l2[1] l1[2]=l2[2] ……..l1[k-1]<l2[k-1] ) then

{

c= l1[1].l1[2] .l1[3].l1[4] ……..l1[k-1].l2[k-1];

}

If c cannot be pruned then

c.tidlist=l1.tidlist l2.tidlist

if( |c.tidlist|/|p|>=min. Sup.) then

Lkp= Lkp {c}

End

End
p
} return k Lk

For shifting partitions right (f(x))
{int array[n]
for (i=n-1;i>=0;i--)
{
array [i]=array[i-1];
}}
For shifting contents in pipeline stages to next stage (f(y))
{int array[m]
for (i=m-1;i>=0;i--)
{
array [i]=array[i-1];
}}

X={pipeline,partition}

F(pipeline);(Now shifting algorithm for pipeline)

F (partition);(shifting of partition to pipeline)



5. DISCUSSION

The efficiency of our algorithm can be highly improved if it is performed on multiprocessor system
and used divide & rule method. In the Apriori algorithm, the support plays a decisive role, but now
the confidence will be in the first place to mine some association rules with very high degree of
confidence. The entire algorithm is divided into three steps: generate optional term sets, filter
optional large term sets, and generate association rules. In the three steps, the key is filtering
candidate large term sets. After a first scan on the database, the initial optional term set is
generated, and after the support is calculated ,the affair data term with low support is found, and
then the confidence of it in the corresponding original database will be calculated through the
method based on probability, if too high, the affair data term set which it belongs to will be filtered
to reduce the number of records in next search; thus in the second scanning database, the
scanning scope will be reduced and the searching time will be shorten, which can improve the
algorithm efficiency by the same way, finally a large optional set will be generated, resulting
association rules, which will be output.

6. CONSLUSION & FUTURE WORK
In this paper we described the data mining approach applied to the data cloud spread across the
globe and connected to the web. We used the Sector/Sphere framework integrated to
association rule based data mining. This enables the application of the association rule
algorithms to the wide range of Cloud services available on the web. We have described a cloud-
based infrastructure designed for data mining large distributed data sets over clusters connected
with high performance wide area networks. Sector/Sphere is open source and available through
Source Forge. The discovery of the association rule is a most successful and most vital duty in
the data mining, is a very active research area in current data mining, its goal is to discover all
frequent modes in the data set, and the current research work carrying on are mostly focused on
the development of effective algorithm. On the basis of in-depth study of the existing data mining
algorithms, in this paper a new data mining algorithm based on association rules is presented.
The algorithm can avoid redundant rules as far as possible; the performance of the algorithm can
be obviously improved when compared with the existing algorithms. We are implementing the
algorithm and comparing our algorithm with other approaches.

7. REFERENCES
[1] Amazon. Amazon Simple Storage Service (Amazon S3). www.amazon.com/s3.
[2] Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung. “The Google File System”. In
SOSP, 2003.
[3] Dhruba Borthaku. “The hadoop distributed file system: Architecture and design”.
retrieved fromlucene.apache.org/hadoop, 2007.

[4] Jeffrey Dean and Sanjay Ghemawat. “MapReduce: Simplifieddata processing on large
clusters”. In OSDI’04: SixthSymposium on Operating System Design andImplementation,
2004.

[5] Amazon Web Services LLC. “Amazon web services developer connection”. retrieved
from developer.amazonwebservices.com on November 1, 2007.



[6] Hillol Kargupta. Proceedings of Next Generation Data Mining2007. Taylor and Francis,
2008.

[7] Ian Foster and Carl Kesselman. “The Grid 2: Blueprint for a New Computing
infrastructure”. Morgan Kaufmann, San Francisco, California, 2004.

[8] Yunhong Gu and Robert L. Grossman. “UDT: UDP-based data transfer for high-speed
wide area networks”. Computer Networks, 51(7):1777—1799, 2007.

[9] I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H Balakrishnana. “Chord: A scalable
peer to peer lookup service for internet applications”. In Proceedings of the ACM
SIGCOMM ’01, pages 149–160, 2001.

[10] Han J , Kamber M. “Data Mining: Concepts and Techniques”. 2/e San Francisco: CA.
Morgan Kaufmann Publishers, an imprint of Elsevier. pp-259-261, 628-640
(2006)

[11] Agrawal R. Imielinski T, Swami. “A Database mining: a performance perspective”. IEEE
Transactions on Knowledge and Data Engineering, Dec.1993,5(6): 914 - 925.

[12] Jim Gray and Alexander S. Szalay. “The world-wide telescope”. Science, vol 293:2037–
2040, 2001.

[13] Yunhong Gu and Robert L. Grossman. “UDT: UDP-based data transfer for high-speed
wide area networks”. Computer Networks, 51(7):1777—1799, 2007.

[14] Robert L. Grossman and Yunhong Gu “Data Mining using high performance data clouds:
Experimental Studies using Sector and Sphere”. Retrieved from
http://sector.sourceforge.net/pub/grossman-gu-ncdm-tr-08-04.pdf.

[15] ZouLi & LiangXu, “ Mining Association Rules in Distributed System”, First International
Workshop on Education Technology and Computer Science,. IEEE, 2009.

[16] A. Rosenthal et al. “Cloud computing: A new business paradigm for Biomedical
information sharing “Journal of Biomedical Informatics, Elsevier ,43 (2010) 342–353 343.

[17] G. Stumme et al. “Web Semantics: Science, Services and Agents on the World Wide
Web”, Journal of WEB Semantics, Elsevier, 4 (2006) 124–143.

[18] John P. Hayes ,”Computer Architecture and Organizaton”,3/e, McGraw-HILL
INTERNATIONAL EDITIONS, Computer Science Series, pp- 275-292 (1998)


Mr. Rizwan Ahmad & Dr. Aasia Khanum

Document Topic Generation in Text Mining by Using Cluster Analysis with
EROCK
Rizwan Ahmad qazirizwan.ahmad@yahoo.com
Computer Engineering Department
National University of Science & Technology
Rawalpindi, 46000, Pakistan

Dr. Aasia Khanum aasia@ceme.nust.edu.pk
Computer Engineering Department
National University of Science & Technology
Rawalpindi, 46000, Pakistan

Abstract

Clustering is useful technique in the field of textual data mining. Cluster analysis divides objects
into meaningful groups based on similarity between objects. Copious material is available from the
World Wide Web (WWW) in response to any user-provided query. It becomes tedious for the user
to manually extract real required information from this material. This paper proposes a scheme to
effectively address this problem with the help of cluster analysis. In particular, the ROCK algorithm
is studied with some modifications. ROCK generates better clusters than other clustering
algorithms for data with categorical attributes. We present an enhanced version of ROCK called
Enhanced ROCK (EROCK) with improved similarity measure as well as storage efficiency.
Evaluation of the proposed algorithm done on standard text documents shows improved
performance.

Keywords: Text Mining, Cluster Analysis, Document Similarity, Topic Generation.

1. INTRODUCTION
A considerably large portion of information present on the World Wide Web (WWW) today is in the form of un-
structured or semi-structured text data bases. The WWW instantaneously delivers huge number of these documents in
response to a user query. However, due to lack of structure, the users are at a loss to manage the information
contained in these documents efficiently. In this context, the importance of data/text mining and knowledge discovery
is increasing in different areas like: telecommunication, credit card services, sales and marketing etc [1]. Text mining is
used to gather meaningful information from text and includes tasks like Text Categorization, Text Clustering, Text
Analysis and Document Summarization. Text Mining examines unstructured textual information in an attempt to
discover structure and implicit meanings within the text.

One main problem in this area of research is regarding organization of document data. This can be achieved by
developing nomenclature or topics to identify different documents. However, assigning topics to documents in a large
collection manually can prove to be an arduous task. We propose a technique to automatically cluster these
documents into the related topics. Clustering is the proven technique for document grouping and categorization based
on the similarity between these documents [1]. Documents within one cluster have high similarity with each another,
but low similarity with documents in other clusters.

Various techniques for accurate clustering have been proposed, e.g. K-MEAN [3, 8], CURE [11, 12], BIRCH [10],
ROCK [1, 2], and many others [1], [3], [10], [11]. K-MEAN clustering algorithm is used to partition objects into clusters
while minimizing sum of distance between objects and their nearest center. CURE (Clustering Using Representation)
represents clusters by using multiple well scattered points called representatives. A constant number ‘c’ of well
scattered points can be chosen from ‘2c’ scattered points for merging two clusters. CURE can detect clusters with non-
spherical shapes and works well with outliers [11, 12]. BIRCH (Balance and Iterative Reducing and Clustering using

International Journal of Computer Science & Security (IJCSS), Volume (4) : Issue (2) 176


Hierarchies) is useful algorithm for data represented in vector space. It also works well with outliers like CURE [10].
However, the traditional clustering algorithms fail while dealing with categorical attributes. As they are based on
distance measure so their merging processing is not accurate in case of categorical data. ROCK (Robust Clustering
Algorithm for Categorical Attributes) gives better quality clusters involving categorical data as compared with other
traditional algorithms. Below we first describe the original ROCK approach and then propose our own enhancements
to ROCK which we call the Enhanced ROCK or EROCK approach.

2. ROCK ALGORITHM
ROCK is an agglomerative hierarchical clustering algorithm [2]. The original algorithm used the Jaccard coefficient for
similarity measure but later on a new technique was introduced according to which two points are considered similar if
they share a large enough number of neighbors. The basic steps of ROCK algorithm are:

1. Obtain a random sample of points from the data set ‘S’
2. Compute the link value for each pair of points using the Jaccard coefficient [2]:
T T
1 2
Sim (T , T ) 
1 2
T T
1 2

3. Maintain a heap (as a sparse matrix) for each cluster’s links
4. Perform an agglomerative hierarchical clustering on data using number of shared objects (as indicated by the
Jaccard coefficient) as clustering criterion.
5. Assign the remaining points to the found cluster
6. Repeat steps 1-5 until the required number of clusters has been found.
ROCK algorithm has following advantages over other clustering algorithms [1]:

1. It works well for the categorical data.
2. Once a document has been added to a specific cluster, it will not be re-assigned to another cluster at the
same level of hierarchy. In other words, document switching across the clusters is avoided using ROCK.
3. It uses the concept of links instead of using distance formula for measuring similarity resulting in more flexible
clustering.
4. It generates better quality clusters than other algorithms.
Limitations of ROCK include the following:
1. ROCK algorithm used sparse matrix for storing cluster links.
2. Sparse matrix takes more space so efficiency suffers adversely.
3. Similarity is calculated by using Jaccard coefficient.
4. Similarity function is dependent on document length.

3. PROPOSED ENHANCEMENTS (EROCK)
EROCK approach includes several enhancements to overcome the limitations of the ROCK algorithm. Here we
discuss these enhancements.

First, ROCK algorithm draws random sample from the database. It then calculates links between the points in the
sample. The purposed approach (EROCK) makes use of entire data base for clustering. Every point in the database is
treated as a separate cluster meaning that every document is treated as a cluster. Then the links between these
clusters are calculated. The clusters with the highest number of links are then merged. This process goes on until the
specified numbers of clusters are formed. So by decomposing the whole database, linkage and topic generation will
become efficient.



Second, ROCK algorithm uses similarity measure based on Jaccard coefficient. We propose cosine measure:

v .v
1 2
CosSim ( v , v )  1 2
v v 1 2

where v 1 and v2 are the term frequency vectors. v1. v2 is the vector dot product defined as:
k

 i  vi vi
1 2

i  1

and |v 1| is defined as:

v 1  v .v 1 2

Cosine similarity is independent of the document length. Due to this property processing becomes efficient. Cosine
similarity has advantages over Euclidean distance while applied on large documents (when documents tends of scale
up), Euclidean will be preferred otherwise.

Third, ROCK uses sparse matrix for link information. The sparse matrix requires more space and long list of
references because of which efficiency suffers adversely. In EROCK adjacency list instead of sparse matrix is
proposed for maintaining link information between neighboring clusters. Adjacency list is a preferred data structure
when data is large and sparse. Adjacency list keeps track of only neighboring documents and utilizes lesser space as
compared to sparse matrix. Besides space efficiency it is easier to find all vertices adjacent to a given vertex in a list.

4. IMPLEMENTATION
4.1 Inputs
The EROCK algorithm requires some initial parameters which are necessary for the whole process. Following are the
major inputs to run the algorithm:

 A directory containing text documents (Corpus).
 Threshold for number of clusters to be formed.
 Threshold value for measuring similarity of documents.
 Threshold value for taking top most frequent words for labeling the folders.

4.2 Document Clustering and Topic Generation Using EROCK Algorithm
Basic steps of EROCK are the same as those of ROCK. For document clustering and topic generation, the text files in
the corpus are first converted into documents. Following are the steps involved in making the clusters, using EROCK
algorithm:

1. Build documents from the text file present in the specified folder.
2. Compute links of every document with every other document using cosine similarity measure [2].
3. Maintain neighbors of each document in an adjacency list structure.
4. After computing links for all documents, each document is treated as a cluster.
5. Extract the best two clusters that will be merged to form one cluster. This decision is made on the basis of
goodness measures. In EROCK, goodness measure defined as the two clusters which have maximum
number of links between them [2]. Let these two clusters be u and v.
6. Now merge the two clusters u and v. Merging of two clusters involve, merging the names of the two clusters,
the documents of two clusters and links of two clusters. This will result in a merged cluster called w.
7. For each cluster x that belongs to the link of w take following steps:

i. Remove clusters u and v from the links of x.



ii. Calculate the link count for w with respect to x.
iii. Add cluster w to the link of x.
iv. Add cluster x to the link of w.
v. Update cluster x in the original cluster list.
vi. Add cluster x to the original cluster list
vii. Repeat step (iv.) until the required number of clusters are formed or there are no two clusters found to
be merged.
viii. After obtaining the final merged cluster list apply labeling process on each. For labeling, the most
frequent word from each document of a cluster is used. Take top most frequent words based on the
threshold value.
The word with high frequency will be treated as the topic or label for a cluster. All related documents will be placed
under one topic. Physically these documents will be put in folders with topics or label as folder name.

4.3 Output:
 A list of clusters labeled properly.
 Each cluster gets converted into a physical folder/directory on the disk and each folder contains the
documents of the respective cluster.

Figure 1: Application Flow

5. EXPERIMENTAL SETUP AND RESULTS
In order to evaluate the effectiveness of our algorithm, we compared the results obtained from ROCK algorithm with
EROCK algorithm on similar text documents (corpus). We run both the algorithms on different corpus sizes i.e. 10, 20,
40, 50, 100, 200. For final algorithm comparison, the size of the corpus was four hundred (400) documents. Initially
stop words and other useless items were removed from the documents in a pre-processing stage. The first step is to
remove common words, for example, a, the, or, and all using a list of stop words. If a word in a document matches a
word in the list, then the word will not be included as part of the query processing [18]. After the generation of
intermediate form, clustering algorithm was applied on it. We report results from two types of analysis: Cluster Analysis
and Label Analysis.



5.1 Cluster Analysis
We analyzed clustering results by varying the desired number of clusters from one (1) to ten (10). For any target
number of clusters, similarity values (threshold) can have the range from 0.1 to 1.0. Figure 2 shows the overall results
obtained from this study. According to the figure, clusters to be obtained are dependent on similarity threshold values.
The inferences gained from the analysis are given as under:
 If the number of clusters to be obtained is equal to the number of documents then similarity factor has no
effect on the clustering results.
 If the number of clusters to be obtained is less than actual number of documents, then the number of clusters
to be obtained depends upon the similarity threshold.
 Increase in the threshold of top frequent word(s) of cluster will increase the size of the cluster label.
 For the dataset which we used for analysis, EROCK discovered almost pure clusters containing relevant
documents with respect to their topics.

Figure 1: Clusters Obtained w.r.t Similarity values

Table 1 shows cluster labeling by similarity threshold values and number of clusters to be obtained. Here we use ten
documents for summarization. If similarity value is 0.1 and number of clusters to be obtained is 1, then only one single
label or topic will be generated and the entire document will be put under it. If similarity value is 0.2 and number of
clusters to be obtained is 2 then two labels will be generated. If similarity value is 1.0 and numbers of clusters to be
obtained are 10 then all the documents will be labeled separately as clusters. It means that labeling or document
topics are mainly dependent on both similarity threshold values and number of clusters to be obtained. Other
scenarios are very obvious from the table.

Similarity Clusters to Folders (Comma Separated)
Threshold be Obtained
0.1 1 ALERTS ALARMS WORM AURORA COST DATA CLASSIFIERS HOST
0.2 2 ALERTS ALARMS WORM1 AURORA COST DATA CLASSIFIERS HOST, WORM
0.3 3 ALARMS, ALERTS, CLASSIFIERS, WORM AURORA COST DATA HOST
0.4 4 ALARMS, ALERTS, AURORA, CLASSIFIERS, DATA, HOST, WORM COST
0.5 5 ALARMS, ALERTS, AURORA, CLASSIFIERS, COST, DATA, HOST, WORM
0.6 6 ALARMS, ALERTS, AURORA, CLASSIFIERS, COST, DATA, HOST, WORM
0.7 7 ALARMS, ALERTS, AURORA, CLASSIFIERS, COST, DATA, HOST, WORM, WORM6

0.8 8 ALARMS, ALERTS, AURORA, CLASSIFIERS, COST, COST8, DATA, HOST, WORM,
0.9 9 WORM7
ALARMS, ALERTS, AURORA, CLASSIFIERS, COST,COST8 DATA, HOST, WORM, WORM7
1.0 10 ALARMS, ALERTS, AURORA, CLASSIFIERS, COST, COST8, DATA, HOST, WORM,
WORM7
Table 1: Cluster analysis results with topics

5.2 Labeling Analysis
Labeling analysis involve the analysis of labels generated for clusters based on similarity threshold values. This
analysis is helpful to check whether the process is accurate or not. Label generation varies as per similarity vales as
shown in Table 1. Following parameters were used for this analysis:
 Number of Documents: 10



 Number of Cluster to be obtained: 4
 Similarity Threshold: 0.3
Top Label Frequency Clusters to be Labels (Comma Separated)
Threshold Obtained
0.3 4 ALARMS, ALERTS, CLASSIFIERS, WORM AURORA
0.5 4 ALARMS, ALERTS, CLASSIFIERS, WORM AURORA COST
0.7 4 ALARMS, ALERTS, CLASSIFIERS, WORM AURORA COST DATA
1.0 4 ALARMS, ALERTS, CLASSIFIERS, WORM AURORA COST DATA HOST

Table 2: Cluster Labeling w.r.t Similarity Value

5.3 Comparison of ROCK & EROCK Performance
It is noteworthy that comparison should be performed on same machine and under same environment. We used the
same machine for comparisons of ROCK & EROCK. It is also necessary that algorithm should also be implemented in
same technology and on same platform. For this we implemented ROCK & EROCK algorithm on same technology and
platform.

Performance results of both algorithms are shown in Figure 2 & Figure 3 and with similarity threshold of 0.1 and 0.5
respectively. We compared the both algorithm with varying sizes of the document (we calculated the number of words
in a document, after removal of stop words). In Figure 2 & Figure 3 we mentioned the document size (No.of Words)
horizontally and time (in second) vertically. It is very clear from both the figures that when document size goes on
increasing, EROCK give good performance as compared with ROCK algorithm.

Figure 2: ROCK vs EROCK with SM=0.1 Figure 3: ROCK vs EROCK with SM=0.5

6. CONCLUSION & FUTURE WORK
In this paper we proposed an efficient way of document topic generation with enhanced version of cosine based
similarity between the pair of categorical data known as clusters. We also proposed and used efficient document
storage technique i.e. adjacency list instead of sparse matrix. By enhancing these two parameters of traditional ROCK
algorithm, we get better results (as shown in Figure 2 & Figure 3). The experimental results obtained from the research
are very encouraging. The outcome of this research shows that by using proposed approach, the cumbersome task of
manually grouping and arranging files becomes very easy. Now user will be able to get relevant information easily
without doing tedious manual activity. Huge information is now available in the form of text documents so
documents/clusters having related information are grouped together and labeled accordingly. Clusters are merged
only if closeness and inter connectivity of items within both clusters are of high significance. Finally it is observed that
EROCK gives good performance for large datasets.



There are many areas in text mining; where one may carry on his/her work to enhance those areas. Out of these, the
labeling of the clusters is a very daunting challenge of this time. No remarkable effort has been made in this regard to
get good result. That is why automatic labeling of the clusters is not so much accurate. A keen and concerted work has
been done to remove this hurdle. It will certainly serve as a lime length for future researchers.

7. REFERENCES
[1] Shaoxu Song and Chunping Li, “Improved ROCK for Text Clustering Using Asymmetric Proximity”, SOFSEM 2006,
LNCS 3831, pp. 501–510, 2006.
[2] Sudipto Guha, Rajeev Rastogi and Kyuseok Shim, “ROCK: A robust clustering algorithm for categorical attributes”.
In: IEEE Internat. Conf. Data Engineering, Sydney, March 1999.
[3] Tapas Kanungo, David M. Mount, Nathan S. Netanyahu, Christine D. Piatko, Ruth Silverman, Angela Y. Wu, “An
Efficient k-Means Clustering Algorithm: Analysis and Implementation”, IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 24, No. 7, July 2002.
[4] Alain Lelu, Martine Cadot, Pascal Cuxac, “Document stream clustering: experimenting an incremental algorithm
and AR-based tools for highlighting dynamic trends.”, International Workshop on Webometrics, Informatics and
Scientometrics & Seventh COLIENT Meeting, France, 2006.
[5] Jiyeon Choo, Rachsuda Jiamthapthaksin, Chun-sheng Chen, Oner Ulvi Celepcikay, Christian Giusti, and Christoph
F. Eick, "MOSAIC: A proximity graph approach for agglomerative clustering," Proceedings 9th International
Conference on Data Warehousing and Knowledge Discovery (DaWaK), Regensbug Germany, September 2007.
[6] S. Salvador, P. Chan, Determining the number of clusters/segments in hierarchical clustering/segmentation
algorithms, Proceedings of the 16th IEE International Conference on Tools with AI, 2004, pp. 576–584.
[7] Murtagh, F., “A Survey of Recent Advances in Hierarchical Clustering Algorithms”, The Computer Journal, 1983.
[8] Huang, Z. (1998). Extensions to the K-means Algorithm for Clustering Large Datasets with Categorical Values.
Data Mining and Knowledge Discovery, 2, p. 283-304.
[9] Huidong Jin , Man-Leung Wong , K. -S. Leung, “Scalable Model-Based Clustering for Large Databases Based on
Data Summarization”, IEEE Transactions on Pattern Analysis and Machine Intelligence, v.27 n.11, p.1710-1719,
November 2005.
[10] Tian Zhang, Raghu Ramakrishan, Miron Livny, “BIRCH: An Efficent Data Clustering Method for Very Large
Databases”.
[11] Linas Baltruns, Juozas Gordevicius, “Implementation of CURE Clustering Algorithm”, February 1, 2005.
[12] Sudipto Guha, Rajeev Rastogi, Kyuseok Shim, “CURE: An Efficient Clustering Algorithm for Large Databases”.
[13] M. Castellano, G. Mastronardi, A. Aprile, and G. Tarricone,”A Web Text Mining Flexible Architecture”, World
Academy of Science, Engineering and Technology 32 2007.
[14] Brigitte Mathiak and Silke Eckstein,” Five Steps to Text Mining in Biomedical Literature”, Proceedings of the
Second European Workshop on Data Mining and Text Mining in Bioinformatics
[15] Ng, R.T. and Han, J. 1994. Efficient and effective clustering methods for spatial data mining. Proceedings of the
20th VLDB Conference, Santiago, Chile, pp. 144–155.
[16] Stan Salvador and Philip Chan, Determining the Number of Clusters/Segments in Hierarchical
Clustering/Segmentation Algorithms, Proc. 16th IEEE Intl. Conf. on Tools with AI, pp. 576–584, 2004.
[17] Sholom Weiss, Brian White, Chid Apte,” Lightweight Document Clustering”, IBM Research Report RC-21684.
[18] Masrah Azrifah Azmi Murad, Trevor Martin,”Similarity-Based Estimation for Document Summarization using Fuzzy
Sets”, IJCSS, Volume (1): Issue (4), pp 1-12.
[19] Sohil Dineshkumar Pandya, Paresh V Virparia, “Testing Various Similarity Metrics and their Permutations with
Clustering Approach in Context Free Data Cleaning”, IJCSS, Volume (3): Issue (5), pp 344-350.


Kamaljit Kaur, Amit Chhabra & Gurvinder Singh

Heuristics Based Genetic Algorithm for Scheduling Static Tasks
in Homogeneous Parallel System

Kamaljit Kaur kamal.aujla86@gmail.com
Department of Computer Science & Engineering,
Guru Nanak Dev University,
Amritsar- 143001, Punjab, India

Amit Chhabra chhabra_amit78@yahoo.com

Gurvinder Singh gsbawa71@yahoo.com

Abstract

Multiprocessor task scheduling is an important and computationally difficult
problem. Multiprocessors have emerged as a powerful computing means for
running real-time applications, especially that a uni-processor system would not
be sufficient enough to execute all the tasks. That computing environment
requires an efficient algorithm to determine when and on which processor a given
task should execute. A task can be partitioned into a group of subtasks and
represented as a DAG (Directed Acyclic Graph), that problem can be stated as
finding a schedule for a DAG to be executed in a parallel multiprocessor system.
The problem of mapping meta-tasks to a machine is shown to be NP-complete.
The NP-complete problem can be solved only using heuristic approach. The
execution time requirements of the applications’ tasks are assumed to be
stochastic. In multiprocessor scheduling problem, a given program is to be
scheduled in a given multiprocessor system such that the program’s execution
time should be minimized. The last job must be completed as early as possible.
Genetic algorithm (GA) is one of the widely used techniques for constrained
optimization. Performance of genetic algorithm can be improved with the
introduction of some knowledge about the scheduling problem represented by
the use of heuristics. In this paper the problem of same execution time or
completion time and same precedence in the homogeneous parallel system is
resolved by using concept of Bottom-level (b-level) or Top-level (t-level). This
combined approach named as heuristics based genetic algorithm (HGA) based
on MET (Minimum execution time)/Min-Min heuristics and b-level or t-level
precedence resolution and is compared with a pure genetic algorithm, min-min
heuristic, MET heuristic and First Come First Serve (FCFS) approach. Results of
the experiments show that the heuristics based genetic algorithm produces much



better results in terms of quality of solutions.

Keywords: DAG, multiprocessor scheduling, genetic algorithm, heuristics.

1. INTRODUCTION
The problem of scheduling parallel tasks onto multiprocessors is to simply apportion a set of
tasks to processors such that the optimal usage of processors and accepted computation time for
scheduling algorithm are obtained [1,2]. The assumption of this paper is based on the
deterministic model, that is, the number of processors, the execution time of tasks, the
relationship among tasks and precedence constraints are known in advance. The precedence
constraints between tasks are represented by a Directed Acyclic Graph (DAG). In addition, the
communication cost between two tasks is considered to be non-negligible and the multiprocessor
system is not diverse and non-preemptive, that is, the processors are homogeneous, and each
processor completes the current task before the new one starts its execution.

The complexity of the scheduling problem is very depended to the DAG, the number of
processors, the execution time of tasks and also the performance criteria which would to be
optimized.

To date, many heuristic methods have been presented to schedule tasks on multiprocessor
systems [5, 9, 10, 11, 16, 18, 19]. Also, there are many studies have been used for task
scheduling based on GA [7, 8, 12, 13, 14, 15, 16, 17, 20, 21, 22, 23]. GA is a problem solving
strategy, based on Darwinian evolution, which has been successfully used for optimization
problems [3, 4].

The aim of this paper is to present a GA to decrease the computation time for finding a
suboptimal schedule.

This paper is divided as follows: In section 2 an overview of the problem is given along with brief
description of the solution methodology. Section 3 provides a more detailed heuristics based
genetic algorithm. Experimental results and performance analysis are provided in section 4 and
conclusion follow in section 5.

2. PROBLEM STATEMENT
In this section, more prescribed multiprocessor scheduling problem and the principles of genetic
algorithms are discussed.

2.1 Task Scheduling Problem
Parallel Multiprocessor system scheduling can be classified into many different classes based on
the characteristics of the tasks to be scheduled, the multiprocessor system and the availability of
information. This paper focus on a deterministic scheduling problem. A deterministic scheduling
problem [1, 2] is one in which all information about the tasks and the relation to each other such
as execution time and precedence relation are known to the scheduling algorithm in advance.
The tasks should be non-preemptive i.e. task execution must be completely done before another
task takes control of the processor, and the processor environment is homogeneous.
Homogeneous of processor means that the processors have same speeds or processing
capabilities.

The main objective is to minimize the total task completion time (execution time + waiting time or
idle time).



The multiprocessor computing consists of a set of m homogeneous processor
P = {pi: i =1, 2, 3…m}
They are fully connected with each other via identical links. Figure 1 shows a fully connected
three parallel system with identical link.

FIGURE 1: A fully connected parallel processor

Consider a directed acyclic task graph G = {V,E} of n nodes. Each node V = {T1, T2,...., Tn} in the
graph represents a task. Aim is to map every task to a set P = {P1,P2, . . . , Pm} of m processors.
Each task Ti has a weight Wi associated with it, which is the amount of time the task takes to
execute on any one of the m homogeneous processors. Each directed edge eij indicates
dependence between the two tasks Ti and Tj that it connects. If there is a path from node Ti to
node Tj in the graph G, then Ti is the predecessor of Tj and Tj is the successor of Ti. The
successor task cannot be executed before all its predecessors have been executed and their
results are available at the processor at which the successor is scheduled to execute. A task is
“ready” to execute on a processor if all of its predecessors have completed execution and their
results are available at the processor on which the task is scheduled to execute. If the next task
to be executed on a processor is not yet ready, the processor remains idle until the task is ready.
The elements set C are the weights of the edges as C = {c : k =1, 2, 3…r} It represents the data
k

communication between the two tasks, if they are scheduled to different processors. But if both
tasks are scheduled to the same processor, then the weight associated to the edge becomes null
[7, 12].

A DAG which has eleven tasks according to their height and their execution time (the time
needed for a task to execute) is shown in Figure 2.

FIGURE 2: An example of a DAG

Tlevel(Ti) is defined to be the length of the longest path in the task graph from an entry task to Ti,
excluding the execution cost of Ti. Symmetrically, blevel(Ti) is the length of the longest path from
Ti to an exit task, including the execution cost of Ti. Formula (2.1) and (2.2) are formal definitions
of tlevel(Ti) and blevel(Ti). Notice that we consider communication costs while calculating values
tlevel and blevel [8].



2.2 Minimum Execution Time (MET)
Minimum Execution Time (MET) assigns each task, in arbitrary order, to the machine with the
best expected execution time for that task, regardless of that machine' availability. The
s
motivation behind MET is to give each task to its best machine [5, 11].
2.3 Min-min Heuristic
Min-min heuristic uses minimum completion time (MCT) as a metric, meaning that the task which
can be completed the earliest is given priority. This heuristic begins with the set U of all
unmapped tasks. Then the set of minimum completion times, M = {min(completion_ time(Ti ,Mj ))
for (1 i n , 1 j m)}, is found. M consists of one entry for each unmapped task. Next, the
task with the overall minimum completion time from M is selected and assigned to the
corresponding machine and the workload of the selected machine will be updated. And finally the
newly mapped task is removed from U and the process repeats until all tasks are mapped (i.e. U
is empty) [5, 11].
2.4 Genetic Algorithms
A genetic algorithm starts with an initial population that evolves through generations and to
reproduce depends on its fitness [3, 4]. In this case, the fitness of an individual is defined as the
difference between its makespan and the one of the individuals having the largest makespan in
the population. The best individual corresponds to the one having the smallest makespan and the
largest fitness.

Next, the operators that compose a genetic algorithm are reviewed. The selection operator allows
the algorithm to take biased decisions favor good individuals when changing generations. For
this, some of the good individuals are replicated, while some of the bad individuals are removed.
As a consequence, after the selection, the population is likely to be dominated by good
individuals. Starting from a population P1, this transformation is implemented iteratively by
generating a new population P2 of the same size as P1.

Genetic algorithms are based on the principles that crossing two individuals can result an
offsprings that are better than both parents and slight mutation of an individual can also generate
a better individual. The crossover takes two individuals of a population as input and generates
two new individuals, by crossing the parents'characteristics. The offsprings keep some of the
characteristics of the parents.

The mutation randomly transforms an individual that was also randomly chosen. It is important to
notice that the size of the different populations is same.

The structure of the algorithm is a loop composed of a selection followed by a sequence of
crossovers and a sequence of mutations. After the crossovers, each individual of the new
population is mutated with some (low) probability. This probability is fixed at the beginning of the
execution and remains constant. The termination condition may be the number of iterations,
execution time, results stability, etc [3, 7, 8, 6].

3. HGA: THE SUGGESTED ALGORITHM
GAs operates through a simple cycle of stages: creation of population strings, evaluation of each
string, selection of the best strings and reproduction to create a new population. The number of
genes and their values in each chromosome depend on the problem specification. In this paper,



the number of genes of each chromosome is equal to the number of the nodes (tasks) in the DAG
and the gene values demonstrate the scheduling priority of the related task to the node (each
chromosome shows a scheduling), where the higher priority means that task must be executed
early. In the basic genetic algorithm the initial population is generated randomly, which can cause
to generate more bad results. To avoid the generation of non-optimal results, heuristic approach
along with precedence resolution can be applied to generate the initial population that gives
better results in terms of quality of solutions.

3.1 Coding of Solutions
For multiprocessor scheduling problem, a schedule is one that satisfies following conditions.
1. The precedence relations among the tasks are satisfied
2. Every task is present and appears only once in the schedule [7, 8].
A schedule can be represented as several lists of computational tasks (fig 3). Each list
corresponds to computational tasks executed on a processor and order of tasks in the list
indicates the order of execution.

FIGURE 3: List Representation of a schedule
3.2 Population Initialization
The next step in the GAs is the creation of the initial population. Number of processors, number of
tasks and population size are needed to generate initial population. Each individual of the initial
population is generated through a minimum execution time or min-min heuristic along with b-level
or t-level precedence resolution to avoid the problem of same execution time or completion time
and same precedence. The problem of same execution time/completion time and precedence
can occur in the homogeneous parallel system as all the processors take same execution time to
execute one task.

The task to be scheduled for each iteration is determined by the following rules:
i. Sort the tasks according to their execution time/completion time in ascending order
according to the minimum execution time (MET)/Min-Min heuristic.
ii. Calculate the bottom-level of each task.
iii. Sort the tasks with the same execution time/completion time and same precedence
according to their bottom-level in descending order.
iv. Assign the tasks to the processors in the order of their bottom-level.

OR

The task to be scheduled for each iteration is determined by the following rules:
i. Sort the tasks according to their execution time/completion time in ascending order
according to the minimum execution time (MET)/Min-Min heuristic.
ii. Calculate the top-level of each task.
iii. Sort the tasks with the same execution time/completion time and same precedence
according to their top-level in ascending order.
iv. Assign the tasks to the processors in the order of their top-level.

The length of all individuals in an initial population is equal to the number of tasks in the DAG.



For example: the initial population of fig. 2 is generated as:

TABLE 1: Priority of execution of tasks based on their execution time, completion time, bottom-level and
top-level.

FIGURE 4: Initial Population of figure 1 using b-level resolution

FIGURE 5: Initial Population of figure 1 using t-level resolution

3.3 Fitness Value
Several optimization criteria can be considered for this problem, certainly the problem is
multiobjective in its general formulation [20]. The elementary criterion is that of minimizing the
makespan, that is, the time when finishes the latest job. A secondary criterion is to minimize the
flowtime that is, minimizing the sum of finalization times of all the jobs. These two criteria are
defined as follows:

and

Fj denotes the time when job j finalizes, Schd is the set of all possible schedules and jobs is the
set of all jobs to be scheduled.
For example fitness value of the initial population is as follows:



FIGURE 6: Assignment of tasks to processors using b-level resolution
The fitness value is calculated in terms of Makespan and Flowtime as discussed above as
Makespan = 24 time units
Flowtime = 3+2+3+4+5+12+11+22+15+24+19 = 120 time units.

FIGURE 7: Assignment of tasks to processors using t-level resolution
The * denotes the communication time and # denotes the waiting time.

3.4 Selection Operator
The design of the fitness function is the basic of selection operation, the design of the fitness
function will directly affect the performance of genetic algorithm. GAs uses selection operator to
select the superior and eliminate the inferior. The individuals are selected according to their
fitness value. Once fitness values have been evaluated for all chromosomes, good chromosomes
can be selected through rotating roulette wheel strategy. This operator generate next generation
by selecting best chromosomes from parents and offspring.

3.5 Crossover Operator
Crossover operator randomly selects two parent chromosomes (chromosomes with higher values
have more chance to be selected) and randomly chooses their crossover points, and mates them
to produce two child (offspring) chromosomes. In this paper one and two point crossover
operators are examined. In one point crossover, the segments to the right of the crossover points
are exchanged to form two offspring as shown in figure 8 (a) and in two point crossover [3] [8],
the middle portions of the crossover points are exchanged to form two offspring as shown in
figure 8 (b).



Randomly selects Parent 1 & 2, crossover point 2

FIGURE 8(a): One Point Crossover

Randomly selects parent 1 & 2, crossover points 1 & 3



FIGURE 8(b): Two Point Crossover

3.6 Mutation
Mutation ensures that the probability of finding the optimal solution is never zero. It also acts as a
safety net to recover good genetic material that may be lost through selection and crossover.
Implementation of two mutation operators is there in HGA. The first one selects two tasks
randomly and swaps their allocation parts. The second one selects a task and alters its allocation
part at random. These operators can always generate feasible offspring, too. Figure 9(a), 9(b),
9(c) & 9(d) demonstrate the mutation operation.

FIGURE 9(a): A Gantt chart before mutation operation

FIGURE 9(b): A Gantt chart after swap mutation operation.
The mutation operation swaps task t6 on processor p1 to task t7 on processor p3.



FIGURE 9(c): A Gantt chart after swap mutation operation.
The mutation operation swaps task t9 on processor p1 to task t8 on processor p3.

FIGURE 9(d): A Gantt chart after swap mutation operation
The mutation operation swaps task t8 on processor p1 to task t11on processor p2 that takes 14
time units to complete.
The procedure of the Suggested Heuristics based Genetic Algorithm is:
Step 1: Setting the parameter
Set the parameter: Read DAG (number of tasks n, number of processors m and comm.
cost), population size pop_size, crossover probability pc, mutation probability pm, and
maximum generation maxgen.
Let generation gen = 0
Step 2: Initialization
Generate pop_ size chromosomes using minimum execution time (MET)/Min-Min
heuristic and b-level/t-level precedence resolutions.
Step 3: Evaluate
Calculate the fitness value of each chromosome
Step 4: Crossover
Perform the crossover operation on the chromosomes selected with probability pc.
Step 5: Mutation
Perform the swap/move mutation on chromosomes selected with probability pm.
Step 6: Selection
Select pop_size chromosomes from the parents and offspring for the next generation.
Step 7: Stop testing
If gen = maxgen, then output best solution and stop
Else gen = gen + 1 and return to step 3

4. EXPERIMENTAL RESULTS & PERFORMANCE ANALYSIS
The final best schedule obtained by applying the suggested algorithm to the DAG of figure 2 onto
the parallel multiprocessor system in figure 1, is shown in figure 10 & 11. The completion time
obtained by heuristics based method using b-level resolution is 14 time units and with t-level
resolution is 16 time units. We also compare the results with FCFS scheduling method, min-min



scheduling method, MET scheduling method and also with the Basic Genetic Algorithm (BGA) [7]
on parallel systems and execution of the schedule are shown in figure 12, 13, 14 & 15.
After applying the suggested heuristics based GA, the best schedule found using b-level
precedence resolution is:
P1: T2 T4 T7 T11
P2: T1 T5 T8 T10
P3: T3 T6 T9

FIGURE 10: A Gantt chart of Suggested Heuristics based Genetic Algorithm using b-level resolution.

After applying the suggested heuristics based GA, the best schedule found using t-level
precedence resolution is:
P1: T2 T4 T7 T11
P2: T1 T5 T10 T8
P3: T3 T6 T9

FIGURE 11: A Gantt chart of Suggested Heuristics based Genetic Algorithm using t-level resolution.

Min-min scheduling policy assigns the tasks to processors p1, p2 & p3 as:
P1: T2 T4 T6 T11
P2: T1 T5 T8 T10
P3: T3 T7 T9

FIGURE 12: A Gantt chart of Min-min scheduler




FCFS scheduling Policy assign the tasks to processors p1, p2 & p3 as:
P1: T1 T4 T7 T10
P2: T2 T5 T8 T11
P3: T3 T6 T9

FIGURE 13: A Gantt chart of FCFS scheduler.

Minimum Execution Time (MET) Scheduling Policy assigns the tasks to processors p1, p2 & p3
as:
P1: T2 T5 T7 T8
P2: T4 T3 T10 T11
P3: T1 T6 T9

FIGURE 14: A Gantt chart of MET scheduler.

After applying the Basic GA, the best schedule found is:
P1: T1 T4 T7 T10
P2: T2 T5 T8 T11
P3: T3 T6 T9

FIGURE 15: A Gantt chart of BGA scheduler.




In Figure 16 (a) & (b), it is clear that HGA can considerably decreases the scheduling time.

(a)

(b)
FIGURE 16: Experimental results for (a) Makespan (b) Flowtime

Performance Analysis
1. Suggested Heuristics based GA using b-level resolution:
Speed up (S): speed up is defined as the completion time on a uniprocessor divided by
completion time on a multiprocessor system.
S = 30/14
= 2.142
Efficiency (E): (S * 100)/ m, where m is the number of processors.
E = (2.142 * 100) / 3 = 71.42 %
2. Suggested Heuristics based GA using t-level resolution
S = 30/16 = 1.875
E = (1.875 * 100) / 3 = 62.5 %
3. Min-min Scheduler:
S = 30/20 = 1.5
E = (1.5 * 100) / 3 = 50 %
4. FCFS Scheduler:
S = 30/21 = 1.428
E = (1.428 * 100) / 3 = 47.61 %
5. MET Scheduler:



S = 30/24 = 1.25
E = (1.25 * 100) / 3 = 41.66 %
6. BGA Scheduler:
S = 30/21 = 1.428
E = (1.428 * 100) / 3= 47.61 %

The performance analysis of various scheduling schemes is shown in figure 17.

FIGURE 17: Performance analysis of min-min, FCFS, BGA, MET, HGA algorithms.

5. CONCLUSION
The task scheduling problem in the distributed systems is known to be NP-hard. The heuristic
algorithms which obtain near-optimal solution in an acceptable interval time are preferred to the
back tracking and the dynamic programming. The genetic algorithm is one of the heuristic
algorithms which have the high capability to solve the complicated problems like the task
scheduling.
In this paper, a new genetic algorithm, named Heuristics based Genetic Algorithm for
Scheduling Static Tasks in Homogeneous parallel System is presented which its population size
and the number of generations depends on the number of tasks. This algorithm tends to minimize
the completion time and increase the throughput of the system. The heuristics based method
found a best solution for assigning the tasks to the homogeneous parallel multiprocessor system.
Experimental results and performance of the heuristics based GA with different precedence
resolution methods is compared with Min-min, MET, FCFS and BGA Scheduling method and
shows the efficiency of 71.42 %. The performance study is based on the best randomly
generated schedule of the suggested GA.

6. REFERENCES
[1] Ishfaq Ahmad, Yu-Kwong Kwok, Min-You Wu, “Analysis, Evaluation, and Comparison of
Algorithms for Scheduling Task Graphs on Parallel Processors”, Proceedings of the 1996
International Symposium on Parallel Architectures, Algorithms and Networks, Page: 207,
1996, ISBN: 0-8186-7460-1, IEEE Computer Society Washington, DC, USA.
[2] Yu-Kwong Kwok and Ishfaq Ahmad, “Static Scheduling Algorithms for Allocating Directed
Task Graphs to Multiprocessors”, ACM Computing Surveys, vol. 31, Issue. 4, December
1999, ISSN: 0360-0300, ACM New York, NY, USA.
[3] D. E. Goldberg, “Genetic algorithms in search, optimization & machine learning”, Addison
Wesley, 1990.



[4] Melanie Mitchell, “An Introduction to Genetic algorithms”, The MIT Press, February 1998.
[5] Tracy D. Braunt, Howard Jay Siegel, Noah Beck, Bin Yao, Richard F. Freund, “A Comparison
Study of Eleven Static Heuristics for Mapping a Class of Independent Tasks onto
Heterogeneous Distributed Computing Systems”, Journal of Parallel and Distributed
Computing, Volume 61, Issue 6, June 2001, Pages: 810-837, ISSN: 0743-7315, Academic
Press, Inc. Orlando, FL, USA.
[6] Ricardo C. Correa, Afonso Ferreira, Pascal Rebreyend, “Scheduling Multiprocessor Tasks
With Genetic Algorithms”, IEEE Transactions on Parallel and Distributed Systems, Vol. 10,
Issue. 8, August 1999, Pages: 825-837, ISSN: 1045-9219, IEEE Press Piscataway, NJ, USA.
[7] Edwin S. H. Hou, Nirwan Ansari, Hong Ren, “A Genetic Algorithm for Multiprocessor
Scheduling”, IEEE Transactions on Parallel and Distributed Systems, vol. 5, Issue. 2,
February2003, Pages: 113-120, ISSN: 1045-9219, IEEE Press Piscataway, NJ, USA.
[8] Amir Masoud Rahmani, Mohammad Ali Vahedi, “A novel Task Scheduling in Multiprocessor
Systems with Genetic Algorithm by using Elitism stepping method”, Science and Research
branch, Tehran, Iran, May 26, 2008.
[9] Martin Grajcar, “Genetic List Scheduling Algorithm for Scheduling and Allocating on a
Loosely Coupled Heterogeneous Multiprocessor System”, Proceedings of the 36th annual
ACM/IEEE Design Automation Conference, New Orleans, Louisiana, United States, Pages:
280 – 285, 1999, ISBN: 1-58133-109-7, ACM New York, NY, USA.
[10] Martin Grajcar, “Strengths and Weaknesses of Genetic List Scheduling for Heterogeneous
Systems”, IEEE Computer Society, Proceedings of the Second International Conference on
Application of Concurrency to System Design, Page: 123, ISBN: 0-7695-1071-X, IEEE
Computer Society Washington, DC, USA, 2001.
[11] Hesam Izakian, Ajith Abraham, Vaclav Snasel, “Comparison of Heuristics for scheduling
Independent Tasks on Heterogeneous Distributed Environments”, Proceedings of the 2009
International Joint Conference on Computational Sciences and Optimization, Volume 01,
Pages: 8-12, 2009, ISBN:978-0-7695-3605-7, IEEE Computer Society Washington, DC,
USA.
[12] Yi-Hsuan Lee and Cheng Chen, “A Modified Genetic Algorithm for Task Scheduling in
Multiprocessor Systems”, Proc. of 6th International Conference Systems and Applications,
pp. 382-387, 1999.
[13] Amir Masoud Rahmani and Mojtaba Rezvani, “A Novel Genetic Algorithm for Static Task
Scheduling in Distributed Systems”, International Journal of Computer Theory and
Engineering, Vol. 1, No. 1, April 2009, 1793-8201.
[14] Michael Rinehart, Vida Kianzad and Shuvra S. Bhattacharyya, “A modular Genetic Algorithm
for Scheduling Task Graphs”, Technical Report UMIACS-TR-2003-66, Institute for Advanced
Computer Studies University of Maryland at College Park, June 2003.
[15] Pai-Chou Wang, W. Korfhage, “Process Scheduling with Genetic Algorithms”, Proceedings of
the 7th IEEE Symposium on Parallel and Distributed Processing, Page: 638, ISBN: 0-8186-
7195-5, October 2005, IEEE Computer Society Washington, DC, USA.
[16] Prof. Sanjay R Sutar, Jyoti P. Sawant, Jyoti R. Jadhav, “Task Scheduling For Multiprocessor
Systems Using Memetic Algorithms”, http://www.comp.brad.ac.uk/het-net/tutorials/P27.pdf
[17] Andrew J. Page and Thomas J. Naughton, “Framework for Task scheduling in
heterogeneous distributed computing using genetic algorithms”, 15th Artificial Intelligence
and Cognitive Science Conference, 2004, Castlebar, Co. Mayo, Ireland, isbn = 1-902277-89-
9 pages = 137-146.
[18] Clayton S. Ferner and Robert G. Babb, “Automatic Choice of Scheduling Heuristics for
Parallel/Distributed Computing”, IOS Press Amsterdam, The Netherlands, Volume 7, Issue 1,
Pages: 47 – 65, January 1999, ISSN:1058-9244.
[19] C.L. McCreary, A.A. Khan, J. Thompson, M.E. McArdle, “A Comparison of Heuristics for
Scheduling DAGs on Multiprocessors”, Eighth International Proceedings on Parallel
Processing Symposium, pages: 446-451, Location: Cancun, ISBN: 0-8186-5602-6, DOI:
10.1109/IPPS.2002.288264, 06 August 2002.
[20] Javier Carretero, Fatos Xhafa, Ajith Abraham, “Genetic Algorithm Based Schedulers for Grid
Computing Systems”, International Journal of Innovative Computing, Information and Control,
ICIC International, Vol.3, No. 6, ISSN 1349-4198, pp. 1053-1071, December 2007.



[21] Annie s. Wu, Han Yu, Shiyuan Jin, Kuo-Chi Lin, and Guy Schiavone, “An Incremental
Genetic Algorithm Approach to Multiprocessor Scheduling”, IEEE Transactions on Parallel
and Distributed Systems, Vol.15, No. 9, On page(s): 824 – 834, ISSN: 1045-9219, INSPEC
Accession Number:8094176, Digital Object Identifier: 10.1109/TPDS.2004.38, 13 September
2004.
[22] Michael Bohler, Frank Moore, Yi Pan, “Improved Multiprocessor Task Scheduling Using
Genetic Algorithms”, Proceedings of the Twelfth International FLAIRS Conference, WPAFB,
OH 45433, American Association for Artificial Intelligence, 1999.
[23] Marin Golub, Suad Kasapovic, “Scheduling Multiprocessor Tasks with Genetic Algorithms”,
OACTA Press, from proceeding (351) Applied Informatics, 2002.
[24] M. Nikravan and M. H. Kashani, “A Genetic Algorithm for Process Scheduling in Distributed
Operating Systems considering Load Balancing”, Proceedings 21st European Conference on
Modelling and Simulation Ivan Zelinka, Zuzana Oplatkova, Alessandra Orsoni, ECMS 2007,
ISBN 978-0-9553018-2-7, ISBN 978-0-9553018-3-4 (CD).
[25] Shuang E Zhou, Yong Liu, Di Jiang, “A Genetic-Annealing Algorithm for Task Scheduling
Based on Precedence Task Duplication”, CIT, Proceedings of the Sixth IEEE International
Conference on Computer and Information Technology, Page: 117, 2006, ISBN: 0-7695-2687-
X, IEEE Computer Society Washington, DC, USA.


Oladipupo O.O. & Oyelade O.J.

Knowledge Discovery from Students’ Result Repository:
Association Rule Mining Approach

Oladipupo O.O. frajooje@yahoo.com
College of Science and Technology, Department
of Computer and Information Sciences
Covenant University
Ota,PMB 1023, Nigeria

Oyelade O.J. ayo2006ola@yahoo.com
College of Science and Technology, Department
of Computer and Information Sciences
Covenant University
Ota,PMB 1023, Nigeria

Abstract

Over the years, several statistical tools have been used to analyze students’
performance from different points of view. This paper presents data mining in
education environment that identifies students’ failure patterns using association
rule mining technique. The identified patterns are analysed to offer a helpful and
constructive recommendations to the academic planners in higher institutions of
learning to enhance their decision making process. This will also aid in the
curriculum structure and modification in order to improve students’ academic
performance and trim down failure rate. The software for mining student failed
courses was developed and the analytical process was described.

Keyword: Association rule mining, Academic performance, Educational data mining, Curriculum,
Students’ Result Repository.

1. INTRODUCTION

Data mining is data analysis methodology used to identify hidden patterns in a large data set. It
has been successfully used in different areas including the educational environment. Educational
data mining is an interesting research area which extracts useful, previously unknown patterns
from educational database for better understanding, improved educational performance and
assessment of the student learning process [1]. It is concerned with developing methods for
exploring the unique types of data that come from educational environment which include
students’ results repository.

Students’ result repository is a large data bank which shows the students raw scores and grades
in different courses they enrolled for during their years of attendance in the institution. Student
performance score is basically determined by the sum total of the continuous assessment and the
examination scores. In most institutions the continuous assessment which includes various
assignments, class tests, group presentations is summed up to weigh 30% of the total score
while the main semester examination is 70%. To differentiate different students’ performances
and scores a set of alphabetic grade is identified to represent the score ranges such as 70-100



as “A”, 60-69 as “B”, 50 to 59 as “C” and 45-49 as “ D” and < 45 as “F”. Any score < 45 is
regarded as a fail performance. This grade representation is different from one higher institution
to another.

From the standpoint of the e-learning scholars, data mining techniques is said to have been
applied to solve different problems in educational environment which includes Students’
classification based on their learning performance; detection of irregular learning behaviors; e-
learning system navigation and interaction optimization; clustering according to similar e-learning
system usage; and systems’ adaptability to students’ requirements and capacities and so on. [2]
The choice of data mining tool is mostly determined by the scope of the problem and the
expected analysis result.

In [3] an approach to classify students in order to predict their final year grade based on the
features extracted from logged data in an educational web-based system was reported. Data
mining classification process was used in conjunction with genetic algorithm to improve the
prediction accuracy. Also, in [4] student data was mined to characterize similar behavior groups
in unstructured collaboration using clustering algorithms. The relationship between students’
university entrance examination results and their success was studied using cluster analysis and
k-means algorithm techniques in [5]. Fuzzy logic concept was not behind in the field of
educational data mining [6,7,8,9], for instance a two-phase fuzzy mining and leaning algorithm
was described in [10], this is an hybrid system of association rule mining apriori algorithm with
fuzzy set theory and inductive learning algorithm to find embedded information that could be fed
back to teachers for refining or reorganizing the teaching materials and test. Association rule
mining technique has also been used in several occasions in solving educational problems and to
perform crucial analysis in the educational environment. This is to enhance educational standards
and management such as investigating the areas of learning recommendation systems, learning
material organization, student assessments, course adaptation to the students’ behaviour and
evaluation of educational web sites [1,11,12 13,14]. In [12] a Test Result Feedback (TRF) model
that analyses the relationships between students’ learning time and the corresponding test results
was introduced. Knowledge Discovery through Data Visualization of Drive Test Data was carried
out in [15]. Genetic algorithm as Ai technique was for data quality mining in [16 ] Association rule
mining was used to mining spartial Gene Expressing [17 ] and to discover patterns from student
online course usage in [14] and it is reported that the discovered patterns from student online
course usage can be used for the refinement of online course. Robertas, in [18] analysed student
academic results for informatics course improvement, rank course topics following their
importance for final course marks based on the strength of the association rules and proposed
which specific course topic should be improved to achieve higher student learning effectiveness
and progress.

In view of the literature, it is observed that different analysis has been done on students’ result
repository but the failed courses in isolation has never been analysed for hidden and important
patterns, which could be of a great importance to academic planners in enhancing their decision
making process and improving student performance. In order to bridge this gap, this paper
presents an analysis of students’ academic failed courses in isolation using association rule
mining. This is to discover the hidden relationships that exist between different students failed
courses in form of rules. The generated rules are analysed to make useful and constructive
recommendations to the academic planners. This promised to enhance academic planner’s
sense of decision making and aid in the curriculum structure and modification which in turn
improve students’ performance and trim down failure rate.

1.1 Association Rule mining

Association rule mining associates one or more attributes of a dataset with another attributes, to
discover hidden and important relationship between the attributes, producing an if-then statement
concerning attribute values in form of rules. (19,20). The formal definition of association rule



mining is : Let I = {i1,i2,…,im} be a set of literals called items and D be a set of transactions where
each transaction T is a set of items such that T I. Associated with each transaction is a unique
identifier, called its TID. We say that a transaction T contains X, a set of some items in I, if X T.
Association rule mining process could be decomposed into two main phases to enhance the
implementation of the algorithm. The phases are:

1. Frequent Item Generation: This is to find all the itemsets that satisfy the minimun
support threshold. The itemsets are called frequent itemsets.
2. Rule Generation: This is to extract all the high confidence rules from the frequent
itemsets found in the first step. These rules are called strong rules.

Over the years different algorithms have been proposed in the literature that implement the two
phases of association rule mining [21]. In this paper the traditional Apriori algorithm is
implemented to generate the hidden patterns from the students’ failed courses dataset which
when analysed will serve as a strong convincing recommendation to academic planning
department in institutions of learning for curriculum structure and modification in other to improve
the students’ performances and minimize failure rate percentage.

1.2 Definition of terms

Association rules: An association rule is an implication expression of the form X Y, where XI,
YI, and X and Y are disjoint itemsets, i.e XY = . The strength of an association rule can be
measured in terms of its support and confidence. The rule X Y holds in the transaction set D
with confidence c and support s, if c% of the transactions in D that contains X also contains Y,
and s% of transactions in D contains XY. Both the antecedent and the consequent of the rule
could have more than one Item. The formal definitions of these two metrics are:

(1)

(2)
Example 1: Consider a rule { CSC111,CSC121}  {CSC211}. If the support count for {CSC111,
CSC121, CSC211} is 2 and the total number of transactions is 5 then, the rule’s
support is 2/5 = 0.4. The rule’s confidence is obtained by dividing the support
count for {CSC 111, CSC121, CSC211} by the support count for {CSC111,
CSC121}. If there are 3 transactions that contain CSC111, CSC121 then, the
confidence for this rule is 2/3 = 0.67. If the minimum rule support is 0.3 and
minimum confidence is 0.5, then, the rule
{ CSC111,CSC121}  {CSC211} is said to be strong, that is; the interestingness of
the rule is high.

1.3 Justification for Support and Confidence measure

Support is an important measure because a rule that has a very low support may occur by
chance. A low support rule in this context is likely to be uninteresting from the academic
perspective because such a failure combination might come accidentally and it might not be
profitable to enhance academic planner decision. Also, confidence on the other hand, measures
the reliability of the inference made by a rule. So, the higher the confidence, the more frequent
the failed courses appear together within the database.



2. METHODOLOGY:

2.1 Development of an Apriori Algorithm

The algorithm starts by collecting all the frequent 1-itemsets in the first pass based on the
minimum support. It uses this set (called L1) to generate the candidate sets to be frequent in the
next pass (called C2) by joininig L1 with itself. Any item that is in C1 and not in L1 is eliminated
from C2. This is achieved by calling a function called ‘apriori-gen’. This reduces the item size
drastically. The algorithm continues in the same way to generate the Ck, of size k from the large
itemsets of k-1, then reduces the candidate set by eliminating all those items in k-1 with support
count less than minimum support. The algorithm terminates when there are no candidates to be
counted in the next pass. Figure 1 shows the general Pseudocode for association rule mining and
Figure 2 shows the traditional apriori algorithm, while figure 3 , shows the algorithm for ‘apriori-
gen’ function called for candidate generation and elimination of non-frequent itemset.

Step 1: Accept the minimum support as minsup and minimum confidence as minconf and
the student failed course as the input data set.
Step 2: Determine the support count for all the item as s (courses under consideration).
Step 3: Select the frequent items; item with s  minsup
Step 4: The set candidate k- item is generated by 1- extension of the large (k-1) itemsets
generated in step3
Step 5: Support for the candidate k-itemsets are generated by a pass over the database.
Step 6; Itemset that do not have minsup are discarded and the remaining itemsets are
called
large k-itemsets.
Step 7 : The process is repeated until no more large item.
Step 8: The interesting rules are determined based on the minimum confidence.

FIGURE 1: General Pseudocode for Association Rule Mining

2.2 Apriori Candidate Generation

The apriori-gen is a function called in line 3 of the algorithm1. It takes as agument , the set
of all large (k-1) itemsets. It returns a superset of the set of all large k-itemsets. The description of
the function is given in algorithm 2.

2.3 Rule generation

After all the frequent itemsets have been generated then the rules are determined. In rule
generation, we do not have to make additional passes over the data set to compute the support
of the candidate rules. All needed is to determine the confidence of each rule by using the
support counts computed during frequent itemsets generation.



Find frequent set Lk − 1.

Join Step.
Ck is generated by joining Lk − 1with itself
Prune Step.
Any (k − 1) – itemset that is not frequent cannot be a subset of a frequent k -
itemset, hence should be removed.
where
(Ck: Candidate itemset of size k)
(Lk: frequent itemset of size k)
FIGURE 2: Frequent itemset generation of the Traditional Apriori Algorithm [21]

Apriori
large 1-itemsets that appear in more than transactions }

while
Generate(Lk − 1)
for transactions
Subset(Ck,t)
for candidates

FIGURE 3: Function for generating candidate itemset [21]

3. Result and Rule analysis

In most literature authors focus on the students’ aggregate performances; Grade Point Average
[12,13,14, 18] and their findings are useful majorly for prediction, which might not really improve
the low capacity students’ performances. In this research the association rule mining analysis
was performed based on students’ failed courses. This identifies hidden relationship between the
failed courses and suggests relevant causes of the failure to improve the low capacity students’
performances. Figure 4 shows a snapshot for association mining process interface. In this work it
is observed that the lower the items minimum support, the larger the candidate generated. This
adversely affects the complexity of the system. For instance, in figure 4, if the item minimum



support is 3 and the rule confidence is 0.5, we have 19 frequent itemsets and 114 rules are
generated. Table 2 show the relationship between the minimum support, minimum confidence
and the generated rule and figure 5 gives the graphical representation.

It was observed that the execution time is also inversely proportional to minimum support, since it
increases as minimum support decreases, which confirmed increase in system complexity and
response time as the minimum support decreases as shown on table 2. With all these
observations it shows that to have a less complex system and a constructive, interesting and
relevant patterns the minimum confidence and support should be large enough to trash out
coincidence patterns. Table 3 also displayed some of the rules generated.

FIGURE 4 :A snapshot for Association rule mining process interface

TABLE 1: Relationship between minimum confidence , minimum support and number of generated rules.

Minimum Item(s) support = 3 Minimum item(s) support = 3 Minimum item(s) support = 3
Average Exe.Time = 6sec Average Exe.Time = 14sec Average Exe.Time = 40sec

Min.Conf. #Rules #of freq. #Rules #of freq. #Rules #of freq. Itemset
Itemset Itemset
50% 6 1 114 19 855 152
60% 6 1 97 19 631 152
70% 6 1 82 19 345 152
80% 6 1 51 19 306 152
90% 2 1 49 19 301 152
100% 2 1 49 19 301 152



FIGURE 3: Graphical representation of effect of minsup, minconf on #of rules

3.1 Rule Analysis

Table 2 shows an instance of the rule generated from the simulation. All the rules with confidence
1, are very strong rules, which implies that if a student failed the determinant (antecedent)
course(s), such student will surely fail the dependent (consequent) course(s). Such rules should
not be overlooked in curriculum structure. Also if the rule support is higher, it means that all the
courses involved are failed together by most of the considered students. From rule number 1 one
can deduce that MAT 122, ECO 111CIS 121 with (s =0.06, c = 0.6). This indicates that, the
probability that every student that fails MAT 122, ECO 111 will also fail CIS 121 is 0.6. This type
of rule is not very strong; in some cases it might be overlooked but notwithstanding, the academic
planner can still take it into consideration. In that case, MAT 122 and CIS 121 should not be taken
in the same semester. This kind of failure can be minimized if one becomes a prerequisite to
another. That is, if a student has not passed MAT 122 they will not be allowed to register for CIS
121. Also, we have from rule 8, a strong rule such that ACC111, BUS211 => CSM211 with (s =
0.1, c= 0.83). ACC 111 is introduction to accounting; BUS 211 is introduction to Business and
CSM 211 Mathematical method 1. The first two courses are compulsory courses for the
Management Information System students. ACC111 is a 100 level first semester course while the
other two are 200 level first semester courses. This implies that a student that fails ACC 111 in
100 level should not be allowed to register for BUS 211 or CSM 211 and if possible, the two, so
as to avoid multiple failure.

With all these observations, if academic planners can make use of the extracted hidden patterns
from students’ failed causes using association rule mining approach, it will surely help in
curriculum re-structuring and also, help in monitoring the students’ ability. This will enable the
academic advisers to guide students properly on courses they should enroll for. This, eventually,
tends to increase the student pass rate.

TABLE 2: An instance of rule generated with support and confidence

RuleNO Rule Rule Support Confidence
1 MAT122,ECO111 => CIS121 0.06 0.6
2 CIS111 => CSC212,ECO111 0.06 0.6
4 CIS121,MAT122 => ECO111 0.06 1
5 ACC111 => CHM111,CIS121 0.06 1
6 BUS211,ECO111 => MAT122 0.06 0.6
7 CIS121,BUS211 => CHM111 0.06 0.75
8 ACC111,BUS211 => CSM211 0.1 0.83



4. Conclusion, Recommendation and Future Work

This study has bridge the gap in educational data analysis and shows the potential of the
association rule mining algorithm for enhancing the effectiveness of academic planners and level
advisers in higher institutions of leaning. The analysis was done using undergraduate students’
result in the department of Computer Science from a university in Nigeria. The department offers
two programmes; Computer Science and Management Information Science. A total number of
30 courses for 100 level and 200 level students are considered as a case study. The analysis
reveals that there is more to students’ failure than the students’ ability. It also reveals some
hidden patterns of students’ failed courses which could serve as bedrock for academic planners
in making academic decisions and an aid in the curriculum re-structuring and modification with a
view to improving students’ performance and reducing failure rate. To adopt this approach a
larger number of students should be considered from the first year to the final year in the
institution. This will surely reveal more interesting patterns. Also, the min. confidence should be of
a higher percentage to be able to have more relevant and constructive rules. In future
applications, in order to improve the comprehensibility and applicability of the association rules, it
will be very useful to also provide an ontology that would describe the content of the courses
which will allow the academic planners to understand better the rules that contain concepts
related to the analysed domain.

References

1. B. Dogan, A. Y. Camurcu. “Association Rule Mining from an Intelligent Tutor” Journal of
Educational Technology Systems Volume 36, Number 4 / 2007-2008, pp 433 – 447, 2008
2. F. Castro, A. Vellido, A. Nebot, and F. Mugica. “Applying Data Mining Techniques to e-
Learning Problems”. Evolution of Teaching and Learning Paradigms in Intelligent
Environment ISBN: 10.1007/978-3-540-71974-8_8 Volume 62, pp 183-221. Springer Berlin
Heidelberg, 2007.
3. B.Minaei-Bidgoli, D. A. Kashy, G. Kortemeyer and, W. F. Punch."Predicting student
performance: an application of data mining methods with the educational web-based
system LON-CAPA" In Proceedings of ASEE/IEEE Frontiers in Education Conference,
Boulder, CO: IEEE, 2003.
4. Talavera, L., and Gaudioso, E. “Mining student data to characterize similar behavior
groups in unstructured collaboration spaces”. In Proceedings of the Arti_cial Intelligence in
Computer Supported Collaborative Learning Workshop at the ECAI ,Valencia, Spain, 2004.
5. Ş. Z. ERDOĞAN, M. TİMOR . “A data mining application in a student database”. Journal of
aeronautics and space technologies ,volume 2 number 2 (53-57) 2005.
6. G.J. Hwang. “A Knowledge-Based System as an Intelligent Learning Advisor on Computer
Networks” Journal of Systems, Man, and Cybernetics Vol. 2 , pp.153-158, 1999.
7. G.J. Hwang, T.C.K. Huang,and C.R. Tseng. “A Group-Decision Approach for
EvaluatingEducational Web Sites”. Computers & Education Vol. 42 pp. 65-86 , 2004.
8. G.J. Hwang, C.R. Judy, C.H. Wu, C.M. Li and G.H. Hwang. “Development of an Intelligent
Management System for Monitoring Educational Web Servers”. In proceedings of the 10th
Pacific Asia Conference on Information Systems, PACIS . 2334-2340, 2004.
9. G.D. Stathacopoulou, M. Grigoriadou. “Neural Network-Based Fuzzy Modeling of the
Student in Intelligent Tutoring Systems”. In proceedings of the International Joint
Conference on Neural Networks. Washington ,3517-3521,1999.



10. C.J. Tsai, S.S. Tseng, and C.Y. Lin. “A Two-Phase Fuzzy Mining and Learning Algorithm
for Adaptive Learning Environment”. In proceedings of the Alexandrov, V.N., et al. (eds.):
International Conference on Computational Science, ICCS 2001. LNCS Vol. 2074.
Springer-Verlag, Berlin Heidelberg New York, 429-438. 2001.
11. S. Encheva , S. Tumin. “ Application of Association Rules for Efficient Learning Work-Flow”
Intelligent Information Processing III , ISBN 978-0-387-44639-4, pp 499-504 published
Springer Boston, 2007.
12. H.H. Hsu, C.H. Chen, W.P. Tai. “Towards Error-Free and Personalized Web-Based
Courses”. In proceedings of the 17th International Conference on Advanced Information
Networking and Applications, AINA’03. March 27-29, Xian, China, 99-104, 2003.
13. P. L. Hsu, R. Lai, C. C. Chiu, C. I. Hsu (2003) “The hybrid of association rule algorithms
and genetic algorithms for tree induction: an example of predicting the student course
performance” [Expert Systems with Applications 25 (2003) 51–62.
14. A.Y.K. Chan, K.O. Chow, and K.S. Cheung. “Online Course Refinement through
Association Rule Mining” Journal of Educational Technology Systems Volume 36,
Number 4 / 2007-2008, pp 433 – 44, 2008.
15. S. Saxena, A. S.Pandya, R. Stone, S. R. and S. Hsu (2009) “Knowledge Discovery through
Data Visualization of Drive Test Data” International Journal of Computer Science and
Security (IJCSS), Volume (3): Issue (6) pp. 559-568.
16. S. Das and B. Saha (2009) “Data Quality Mining using Genetic Algorithm” International
Journal of Computer Science and Security, (IJCSS) Volume (3) : Issue (2) pp. 105-112
17. M.Anandhavalli , M.K.Ghose and K.Gauthaman(2009) “Mining Spatial Gene Expression
Data Using Association Rules”. International Journal of Computer Science and Security,
(IJCSS) Volume (3) : Issue (5) pp. 351-357
18. R. Damaševicius. “Analysis of Academic Results for Informatics Course Improvement
Using Association Rule Mining”. Information Systems Development Towards a Service
Provision Society. ISBN 978-0-387-84809-9 (Print) 978-0-387-84810-5 (Online) pp 357-
363, published by Springer US, 2009.
19. Ceglar, J.F Roddick. “Association mining”. ACM Computing Surveys, 38:2, pp. 1-42, 2006
20. S. Kotsiantis, , D. Kanellopoulos. “Association Rules Mining” A Recent Overview.GESTS
Int. Transactions on Computer Science and Engineering, Vol. 32 (1), pp. 71-82, 2006.

21. H. Jochen, G. Ulrich and N. Gholamreza . “Algorithms for Association Rule Mining – A
General Survey and Comparison”. SIGKDD Exploration, Vol.2, Issue 1, pp 58-64. ACM,
2000.


J.Nafeesa Begum, K.Kumar, Dr.V.Sumathy

Multilevel Access Control in a MANET for a Defense Messaging
system using Elliptic Curve Cryptography

J. Nafeesa Begum nafeesa_jeddy@yahoo.com
Sr.Lecturer/CSE
Government College of Engineering,Bargur
Krishnagiri District,Tamil Nadu,India

K.Kumar pkk_kumar@yahoo.com
Lecturer/CSE
Government College of Engineering,Bargur
Krishnagiri District,Tamil Nadu,India

Dr.V.Sumathy sumathy_gct2001@yahoo.co.in
AssistantProfessor/ECE
Government College of Technology,Coimbatorer
Coimbatore District,Tamil Nadu,India

Abstract

The trend of the Civilian society has moved from the industrial age focus on
automation and scale towards information based on computing and
communication. Today’s Warfare is also moving towards an information age
paradigm based on information sharing, situational awareness, and distributed
points of intelligence, command and control. A widely-networked fighting force is
better able to share information about tactical situations that may be
geographically widespread, asymmetric, and rapidly changing. Commanders
must be able to better assess situations across broad theaters, with extensive
data, voice, and especially video feeds as strategic inputs. Thus, network-centric
warfare improves effectiveness at both the tactical "point of the spear" and in the
achievement of broader strategic goals. Broadly disseminated knowledge assets
enable fighting forces that must self-synchronize, even as they physically
disperse to address dynamic battlefield conditions. The speed of decision has
increased and command decisions must be rapidly relayed and implemented, to
improve battlefield outcomes. Multilevel access control in a MANET for a
Defense messaging system is used to have the command decisions relayed to
all people who are active in the group and also to all people who have been
identified as higher in the hierarchy instead of sending one to one messages to
each individual.. The system developed is secure, multi site and allows for global
communication using the inherent properties of Elliptic Curve cryptography .
Elliptic Curve cryptography provides a greater security with less bit size and it is
fast when compared to other schemes. The implementation suggests that it is a
secure system which occupies fewer bits and can be used for low power devices.

Keywords: Defense messaging system,Elliptic Curve cryptography, Encryption , Global Information
Sharing , Secure system.



1. INTRODUCTION
Information superiority has become as important in today's battlefield as air superiority was in the
past in increasing mission effectiveness. Information superiority has become critical as needs of
both war fighters and commanders have broadened to include real-time video, high-speed data,
and voice. Data and intelligence sources include terrestrial forces and sensors, satellites, UAVs
(Unmanned Aerial Vehicles), and a wide variety of centralized and distributed information
assets.[7,8] The vast majority of these information assets, command, communications, and
control must be delivered wirelessly, with seamless connections to wired networks for intelligence
resources and other data. Further, these wireless technologies must support data, voice, and
increasingly, video traffic flows. In the network-centric warfare environment, mobility implies more
than just the motion of individuals and vehicles in relation to one another and to other fixed
locations.

Figure:1 MANET In a military warfare

In such a mobile adhoc network a Defense messaging system takes a message and forwards it
to the intending recipients or parties based on the message criteria for immediate action. It
enables the top defense personnel to issue commands using messages to guard the country
against threats of terrorists, anti socials and Intruders. The system developed has the following
capabilities
 Encryption of the data stream
 Strong user authentication
 Prevention of interaction with external and undesirable applications
 Platform independent
 Central administration and logging

As a multilevel access control systems, the system developed has the following features

 Hiding of hierarchy and receivers
 Authentication of receivers
 Dynamics at message level, class level and user level by the server

The main advantage of ECC compared to other schemes is that it offers equal security with a
smaller key size and thus reduces processing overhead and can be used in Tiny devices also.
The rest of the paper is as follows. Section 2 gives the related work Section 3 describes Elliptic
Curve Cryptography. Section 4 explains our system development, algorithms and dynamics.
Section 5 gives the implementation results, Section 6 does the Performance Analysis and
Section 7 concludes the paper.



2.RELATED WORK

The first multi level access solution was proposed by Akl et al.[1] in 1983 and followed by many
others [8,9,10,11,13,14,15,16 ,12 ]. These schemes basically rely on a one-way function so that a
node v can easily compute v's descendants' keys whereas v's key is computationally difficult to
compute by v's descendant nodes. Moreover, many existing schemes have some of the following
problems: (1) Some schemes were found with security flaws shortly after they were proposed
[3,4] (2) Some schemes cannot support for reconfiguration of a hierarchy [18,19 ]; (3) Some
schemes require access hierarchy to be in a certain form so that it must be a tree or a DAG with
only one root; and (4) Member revocation is one of the most difficult processes in cryptographic
schemes, therefore, it is important to address this problem so that the revocation process can be
dealt with efficiently. In this paper, we propose a new scheme based on elliptic curve
cryptography for secret messaging which has the suitable characteristics. Unlike many existing
schemes based on one-way functions, our scheme is based on a secret sharing method which
makes the scheme unconditionally secure [17, 20] .Elliptic curve cryptography address the issue
of saving the power due to the use of less number of bits for secure transmission [2]. In our
previous work we have used [5,6] Elliptic curves used for efficient group key management. In this
paper we have extended it to include multilevel access control. Multilevel access control using
elliptic curve cryptography is a new research area under deployment and we have used it in the
defense messaging system so that higher group members can see the messages relayed to
lower group members.

3. ELLIPTIC CURVE CRYPTOGRAPHY
3.1 Basics
Elliptic curves are named so as they appear to be similar to the equation defining the roots of an
ellipse. They are equations containing two variables and coefficients where the elements are in a
finite field ( Zp ) . The elliptic equation is of the form y2 =x3 +ax + b. The coefficients a,b should
satisfy the condition 4a3 -27b2 ≠ 0 so that there are no repeated factors. For given values of a and
b , the elliptic curve consists of positive and negative values of y for each value of x. A special
point O which acts as an identity is used. The following addition rules are used in elliptic curve
arithmetic.

1. P + O = O +P = P for all P belongs to Zp
2. If P = ( x, y) Є E (Zp) then ( x, y) +( x, -y) = o and is called the negative of P and –P is a
point on the curve.
3. Let P =(x1,y1) Є E (Zp) Q =(x2,y2) Є E (Zp) where P≠Q then P + Q is =( x3,y3)
x3 = ( λ 2 – x1 - x2 )mod P
y3 = (λ ( x1-x3) –y1 ) mod P and
λ =(( y2- y1) / ( x2-x1)) mod P if P ≠ Q
λ = ((3x12 +a )/ 2y1 ) mod P if P = Q
4. Multiplication is defined by repeated addition

In case of a finite group Ep(a, b) the number of points on the elliptic curve is bounded by P+1 -2
√P ≤ N ≤ P + √P so that for a large P ,the number of elements is approximately equal to P.

3.2 Algorithm For Elliptic Curve Cryptography
Step1: Decide on the elliptic curve E .The Elliptic curve should have two coefficients a,b such that
3 2
4a -27b ≠ 0 and p a prime number.
Step2: For the elliptic curve equation apply values of x from 1 to p-1 and generate y values.
Step3: Find the Quadratic residues to avoid repetition in mod values.
Step4: Collect all the points on the elliptic curve.



Step5 : Use one point called the base point as the generator using which scalar multiplications
are performed and generate multiples of the generator by applying the ECC arithmetic rules.
Step 6: For a same elliptic curve, by choosing a different generator point we can obtain a different
encryption values.

3.3. Example

Points in Elliptic Curve :The Points in E211(0,-4) are found using steps 1 to 5.

1,29 1,182 2,2 2,209 5,11 5,200 6,1 6,210 12,6 12,205
13,100 13,111 14,29 14,182 16,100 16,111 17,30 17,181 19,37 19,174
20,20 20,191 21,87 21,124 23,66 23,145 24,59 24,152 26,103 26,108

Table: 1 Some Points in E211(0,-4)

Using (2,2) as Generator point ,we get the multiple point scalar multiplication of generator from 0
to Infinite Limit. Let 1 G = ( 2 , 2 ) to generate 2 G we perform scalar Multiplications G + G =
2G and using the formulae

x3 = λ 2 – x1 - x2,y3 = λ ( x1-x3) –y1 and λ =( y2- y1) / ( x2-x1) if P ≠ Q
λ = 3x12 +a / 2y1 if P = Q
= 12 /4 = 3,
x3 = 9 – 2 - 2 = 5,
y3=3 (2-5 ) - 2=3 ( -3 ) -2 = -11 mod 211 = 200

ANS:5,200

3.4.Generation of Points using Scalar Multiplication

To get 3G we add G (2,2) to 2 G (5,200) we get different P values as shown below

2,2 5,200 129,56 159,114 153,108 125,152 179,199 174,163 111,145 75,90

168,6 155,96 21,87 201,85 28,2 181,209 150,85 198,139 161,142 54,138

70,200
27,30 84,210 87,50 192,201 69,20 51,136 182,100 64,194 29,139

Table:2 The Scalar Multiplication points for E211(0,-4) at G(2,2)

This procedure is used for generating different elliptic curves. In fact we can use the same elliptic
curve for all classes by changing the generator value. For the above points we can generate 240
different sets each containing 240 points for encryption.

4.Elliptic Curve Cryptography for defense messaging system

4.1.System Overview
Multilevel Access Control in Manet for a defense messaging system is useful for military
organizations which have a hierarchical structure. For example in the Indian Military System the
following hierarchy exists.



CHIEF OF ARMY

ARMY COMMANDER

LIEUTANANT GENERAL

MAJOR GENERAL

BRIGADIER

COLONEL

LIEUTANANT COLONEL

MAJOR

CAPTAIN

LIEUTANANT

Figure 2 : Military hierarchy

In such a type of system, messages sent to a lower class should be known to the active
members of lower class and also to all active members of the higher class. It is not only essential
to maintain the access control but the data should be hidden as well. Elliptic curve cryptography
technique is used .

There are many messages to be sent to different parties. The server inserts new data streams
according to the classification . The messages are encrypted using ECC according to the access
allowed for each user and ,the data is sent . Consider the following set of messages.

Class Category of Data Streams
Confiden Field Messages Terror Climate Warning
tial Messages
Troops × √ × ×
Air Wing × × × √
NSGS × × √ ×
Lieutenants √ √ × ×
Table:3 Example Showing Message classifications

All the users of defense messaging system need to register themselves and get authenticated
by the server. Once the registration process is over the user when joins receives the message
he is entitled to receive .Only authenticated users are able to view the message content as the
message remains unintelligible to people who do not belong to that elliptic curve. Different
Elliptic curves identify different class of users .the servers contribution and the users contribution
are used for finding the Group keys.



Figure 3 :System Overview

L1,L2,L3 belongs to lieutenant group ..N1 is National Security Guard ,T1,T2 are troops and A1 is
a Air wing Officer.

4.2 The Proposed Scheme

The following methodology is used for developing a multilevel access control solution in a Manet
for a defense messaging system using Elliptic Curve Cryptography.

Step1: The server or the central authority creates different classes with different service
requirements. The system is dynamic and any hierarchy can be created or modified. The classes
are arranged as a Tree based structure. A class can have any number of descendent classes at
the next level and a class can have any number of users .
Step2 : Each class maintains a list of its ancestor and descendent classes.
Step 3: Each class is associated with an elliptic curve E : y2 = ( x3 + ax+ b) mod p over Zp and
G = E(Zp)is a Generating point.
Step 4: Users generate a random number as their contribution value and join their classes. The
Shared secret key for each class is calculated whenever user joins or leaves .
Step 5: One user in each class is designated as the class controller . The user joining last is
usually given the role of the class controller .
Step 6: The class controller of each class initiates the process of changing the group key
Whenever users join or leave . Hence forward secrecy and backward secrecy is strictly
maintained and only active members of the particular class will be able to view the message
transfer of that particular class.
Step 7: Whenever the class controller leaves , the user who joined the class before the class
controller now becomes the class controller. This is based on the assumption that a person who
joins late will leave late which is many times true. The group key is changed in this case also so
that the old class controller should not be able to decrypt the message.
Step 8: The group key is sent to all the ancestor class controller nodes by the class controller.
Step 9: The messages are also sent to the ancestor class controllers.



Step10. The messages received by the ancestor class controller will be decrypted by sender’s
group key and again encrypted with their own group key and sent to all users of that class .The
users will use only their Group key to decrypt ,so users need to remember only their group key .
A user who is also a class controller should remember his group key and also the group key of all
its descendent classes.
Step 11: This system is efficient because all dynamics of a class are local to that class only and
only one message is sent to the ancestor class controllers .
Step 12: All users of a higher level class are not disturbed by the dynamics of the lower class and
also they receive the messages from their class controller.
Step 13: The Central Authority is less burdened as all the message transfers are local among the
class controller.

4.3 System Dynamics

1. Message stream Dynamics: New data streams are encrypted and send by the server.
There is a facility to add new items as and when the need arises for the defense
messaging system.
2. Class Dynamics: New Classes can be added by using a new curve and what messages
should be included for the new classes can be set
3. User Dynamics: User may join or leave .Whenever a user joins, he is authenticated and
group key is calculated. Whenever the user leaves the group key of that particular class
gets changed .Hence forward secrecy and backward secrecy are maintained.

4.3 ECC Key Exchange Mechanism :

The key exchange is an analogue of the Diffie Hellmann scheme. The server uses a different
random number as the contribution for each of the curves. The Elliptic curves are decided before
hand and each class user comes in a separate elliptic Curve. The Group keys are used for
getting the messages encrypted. It is found to be very secure.
Users joining the Troop class is shown below

The Server Joins

User Id: Server Private Key (nA) = 47568
Public key (A) = gnA = (nA mod p) G = (47568 mod 241) G
= 91 G = (206,121)
Troop User1 Joins
User Id: TUser1 Private Key (nB) = 13525
nB
Public key (B) = g = (nB mod p) G = (13525 mod 241) G
= 29 G = (29,139)

Finding the Group key after One Troop user and Server joined the group
Server Calculates the Group key
nB
Server will get g from TUser1 i.e. (29,139) yields 29
nAnB
Shared key = g
= ((47568*29)mod 241) G
= 229 G
= (155,115)
Troop User1 Calculates the Group key
TUser1 will get gnA from Server i.e. (206,121) yields 91
Shared key = gnAnB



= ((13525*91)mod 241) G
= 229 G
= (155,115)

Troop User2 Joins the Group
User Id: TUser2 Private Key (nC) = 82910
Public key (C) = gnC = (nC mod p) G = (82910 mod 241) G
= 6 G = (125,152)

Finding the Group key after the Second Troop User joined the Group
The new TUser2 act as a Group controller.
nBnC nAnC
TUser2 computes g ,g
nA
g = (206,121) yields 91
gnAnC = (91*82910 mod 241) G = 64 G = (147,97)
gnB = (29,139) yields 29
nBnC
g = (29*82910 mod 241) G = 174 G = (131,84)
nBnC nAnC
Sends the g Value to Server and g Value to TUser1.

Finding the Group key after three users joined the group

Server Calculates the Group key
server will get gnBnC from TUser2 (GC) i.e. (131,84) yields 174
Shared key = gnAnBnC
= ((47568*174)mod 241) G
= 169 G
= (120,31)
TUser1 Calculates the Group key
TUser1 will get gnAnC from TUser2 (GC) i.e. (147,97) yields 64
= ((13525*64)mod 241) G
= 169 G
= (120,31)
TUser2 Calculates the Group key
gnAnB i.e. (155,115) yields 229
= ((82910*229)mod 241) G
= 169 G
= (120,31)

User Leave from the Group
Let the TUser2 be leave. Then the user sends message to all users that it is leaving. All the users
remove the leaving user from the user list. The group controller changes its key value and
computes the new group key.
Group controller New Private Key = 43297.
The group controller recalculates the following values:
gnAnB= (155,115) yields 229
nB nA
Sends the g Value to Server, g Value to TUser1.
Using the shares the Group keys are calculated

Message Encryption

Message: Enter nestFire
Random Number KA : 202
Cipher Text Pc =( KAG, PM+KASK)



KAG = 202 G
= 202 mod 241 G
= (50,57)
Cipher text: 50:57:

1:182:164:197:203:180:1:182:172:50:136:11:164:197:
1:182:160:8:203:180:114:113:185:199:172:50:1:182:

PM+KASK
PM
e101 (1,182)
n110 (164,197)
t116 (203,180)
e101 (1,182)
r114 (172,50)
(Space)32 (136,11)
n110 (164,197)
e101 (1,182)
s115 (160,8)
t116 (203,180)
f102 (114,113)
i105 (185,199)
r114 (172,50)
e101 (1,182)
KASK = (202 * 18) mod 241
= 21
PM+KASK
e= (101+21) mod 241
= 122 G
= (156,10)
n=(110+21) mod 241
=131 G
=(164,14)
t = (116+21) mod 241
= 137 G
= (163,161)
e= (101+21) mod 241
= 122 G
= (156,10)
r = (114+21) mod 241
= 135 G
= (77,145)
= (32+21) mod 241 (for space)
= 53 G
= (99,180)
n =(110+21) mod 241
=131 G
=(164,14)
e = (101+21) mod 241
= 122 G
= (156,10)
s = (115+21) mod 241
= 136 G
= (185,12)
t = (116+21) mod 241
= 137 G = (163,161)



f = (102+21) mod 241
= 123 G
= (104,190)
i = (105+21) mod 241
= 126 G
= (160,203)
r = (114+21) mod 241
= 135 G
= (77,145)
e = (101+21) mod 241
= 122 G
= (156,10)

Message Decryption

The Group Key (Sk) is (198,139)
Random no. Chosen can be found by the KAG Value (50,57).

From (50,57) we may trace the value KA as 202.

(1)PM+KASK = (156,10)

PM = (156,10) - KASK
= (156,10) -21
= 122 – 21
= 101 G => (1,182)  e
(2) PM+KASK = (164,14)

PM = (164,14) - KASK
= (164,14) -21
= 131 – 21
= 110 G => (164,197)  n

(3) PM+KASK = (163,161)

PM = (163,161)- KASK
= (163,161) -21
= 137 – 21
= 116G => (203,180)  t

(4) PM+KASK = (156,10)

PM = (156,10) - KASK
= (156,10) -21
= 122 – 21
= 101 G => (1,182)  e

(5) PM+KASK = (77,145)

PM = (77,145) - KASK
= (77,145) -21

= 135 – 21
= 114 G => (172,50)  r



(6) PM+KASK = (99,180)
PM = (99,180) - KASK
= (99,180) -21
= 53 – 21
= 32 G => (136,11)  (Space)
(7) PM+KASK = (164,14)
PM = (164,14) - KASK
= (164,14) -21
= 131 – 21
= 110 G => (164,197)  n
(8) PM+KASK = (156,10)
PM = (156,10) - KASK
= (156,10) -21
= 122 – 21
= 101 G => (1,182)  e
(9) PM+KASK = (185,12)
PM = (185,12) - KASK
= (185,12) -21
= 136 – 21
= 115 G (160,8)  s
(10) PM+KASK = (163,161)
PM = (163,161)- KASK
= (163,161) -21
= 137 – 21
= 116G => (203,180)  t
(11) PM+KASK = (104,190)
PM = (104,190) - KASK
= (104,190) -21
= 123 – 21
= 102 G => (114,113)  f
(12) PM+KASK = (160,203)
PM = (160,203) - KASK
= (160,203) -21
= 126– 21
= 105 G => (185,199)  i
(13) PM+KASK = (77,145)

PM = (77,145) - KASK
= (77,145) -21
= 135 – 21
= 114 G => (172,50)  r
(14) PM+KASK = (156,10)
PM = (156,10) - KASK
= (156,10) -21
= 122 – 21
= 101 G => (1,182)  e

4.4.Elliptic Curves used

Troops Class:
y2 =x3 -4 mod 211 at G( 2,2)

NSG Class: (National Security Guards)

y2 =x3 +8x -2 mod 337 at G( 0,311)



Lieutenants:
2 3
y =x +7x +5 mod 563 at G( 1,442)

AIR Wings:
y2 =x3 +5x -8 mod 823 at G( 3,597)

4.5.Example Message:

Sent....3:Leutanats:TerroristInformation:440:400:487:137:493:111:355:172:325:238:54:289:325:2
38:493:111:215:360:16:73:505:466:505:466:538:236:505:466:293:20:560:148:478:249:378:490:1
51:196:408:337:538:236:505:466:552:448:558:532:478:249:293:20:538:236:151:196:493:111:91:
401:325:325:325:325:493:111:91:401:154:529:493:111:347:141:347:141:2:468:493:111:91:401:3
47:141:154:529:493:111:487:137:347:141:347:141:493:111:487:137:110:102:493:111:151:196:3
4:407:84:311:84:311:493:111:151:196:34:407:84:311:84:311:493:111:151:196:34:407:84:311:84:
311:493:111:151:196:34:407:84:311:84:311:493:111:347:141:325:325:341:438:493:111:487:137:
487:137:154:529:493:111:151:196:34:407:84:311:84:311:493:111:151:196:34:407:84:311:84:311
:493:111:347:141:2:468:329:421:493:111:91:401:347:141:154:529:493:111:154:529:493:111:487
:137:91:401:91:401:493:111:215:203:493:111:91:401:154:529:110:102:493:111:341:438:325:325
:493:111:347:141:347:141:341:438:493:111:487:137:325:325:493:111:347:141:110:102:110:102:
493:111:347:141:2:468:329:421:493:111:91:401:347:141:154:529:493:111:347:141:110:102:215:
203:493:111:487:137:487:137:91:401:493:111:347:141:2:468:329:421:493:111:91:401:347:141:1
54:529:493:111:347:141:341:438:487:137:493:111:347:141:347:141:325:325:493:111:91:401:32
5:325:2:468:493:111:347:141:110:102:347:141:493:111:341:438:325:325:493:111:347:141:347:1
41:341:438:493:111:347:141:329:421:325:325:493:111:487:137:347:141:329:421:493:111:215:2
03:493:111:91:401:154:529:110:102:493:111:341:438:325:325:493:111:347:141:347:141:341:43
8:493:111:91:401:325:325:325:325:493:111:487:137:347:141:341:438:493:111:347:141:110:102:
347:141:493:111:91:401:110:102:91:401:493:111:347:141:110:102:215:203:493:111:487:137:48
7:137:91:401:493:111:341:438:325:325:493:111:347:141:347:141:341:438:493:111:91:401:215:2
03:91:401:493:111:91:401:154:529:341:438:493:111:347:141:329:421:325:325:493:111:487:137:
347:141:329:421:493:111:91:401:341:438:341:438:493:111:347:141:347:141:341:438:493:111:3
47:141:329:421:325:325:493:111:487:137:347:141:329:421:493:111:347:141:329:421:325:325:4
93:111:487:137:347:141:329:421:493:111:151:196:34:407:84:311:84:311:493:111:151:196:34:40
7:84:311:84:311

5 IMPLEMENTATION RESULTS:
The system was developed in Java net beans and run on a network. Some sample output
screens are shown

Figure:4 Main Page



Figure:5 Class Join

Figure:6 Class Delete

Figure:7 User Join



Figure :8 Message Received Encrypted and decrypted

Figure :9 Message Received Encrypted And Decrypted by lieutenant class

6. PERFORMANCE ANALYSIS:
6.1. Security :
The Security of ECC is due to the discrete logarithm problem over the points on the elliptic curve.
Cryptanalysis involves determining x given Q and P where P is a point on the elliptic curve and Q
= x P that is P added to itself x times. The best known algorithm to break the elliptic curve points
is the pollard – rho algorithm which is a fully exponential algorithm and difficult o solve.. Forward
and Backward secrecy are maintained as each session is considered as a separate session. In
this section we discuss some attacks and prove that our scheme is secure and feasible. Consider
the Figure 10

Figure :10 Illustration for security



Attack 1: Contrary Attacks
Assuming that E1 (lower privileged user) wants to crack the secret key of B1 ( Higher Privileged
User). It is not possible to decrypt any messages as the hierarchy does not provide secret keys to
descendent classes. Without knowing the secret key it is impossible to see the message.

Attack 2: Interior Collecting Attacks
If a lower level User has many ancestors and if it negotiates with one parent also by knowing the
key as there is no relation parameter between any of the ancestor nodes it is not possible to
derive the key.

Attack 3: Exterior Collecting Attack
If an attacker is outside the system, it means no idea about what elliptic curve or generator point
is being used is known and hence more difficult to attack.

Attack 4: Collaborative Attacks
We assume that if there is a higher privileged user belonging to class B as in figure 10 and there
are two descendant classes D and E. Users of D and E cannot perform a collaborative attack
as the secret key of any class is calculated only from the contribution of the respective users of
the class. If they compromise one user belonging to the higher privileged group to know the
key also, there is no communication possible as the class controllers who control the keying
process also change dynamically .

Attack 5: Sibling Attacks
Classes who have same parent also cannot crack the key of a sibling class due to the absence of
any related parameters among them.

To maintain the secure structure the following things are necessary.
1. The immediate parents should be loyal and the descendant list should be updated.
2. The Class Controller of a Leaving /Changing Class from the tree hierarchy should update their
ancestor list.
3. Recalculation of shared secret key by the leaving /changing class should be done by selecting
a random value for finding a new group key.

The third point is very Important and the execution of this prevents disloyal ancestors also from
finding the key of a left descendant class even if it no longer comes under their control. The most
important feature is that only the key is being transferred and only the authorized entities have
idea about which elliptic curve is used or generator points that are used. This is a very big
advantage as even though an adversary comes across the key pair he may not know the elliptic
curve and the generator. Even users inside the system may decrypt the information but they
may not be aware about the mechanism that takes place.

6.1.1 Enhanced Security Framework:
The keys are always transmitted as plain text but we have justified that even with the key
any attacker will not be able to first of all receive the messages and even if they somehow receive
the message and the key , they will not be able to decrypt the information as they are unaware
about the elliptic curve. If security needs to be enhanced the following framework can be included

1. All first level ancestor – descendant pairs use Diffie- Hell man key exchange and generate a
key which can be used for encrypting and decrypting the actual key.
2. This key can again be combined with other descendant classes. For e.g., in our sample
hierarchy first we perform key exchanges and get the shared secret keys SSK BA, SSK CA and
then again use classes D and E and generate SSK DBA ,SSK EBA we can use the resultant keys
for encrypting and decrypting the original keys.
4. The above scheme prevents repeated computations



6.2. Memory Cost:

Using ECC approach consumes very less memory when compared to RSA and DES. ECC
based approach takes very less memory even the members get increased.

6.3.Communication Costs:

Using other schemes consumes more bandwidth. The Communication and computation of tree
based ECDH depends on trees height, balance of key tree, location of joining tree, and leaving
nodes. But our approach depends on the number of member in the subgroup, number of Group
Controller, and height of tree. So the amount spend on communication is very much less when
compared to CRTDH and RSA based scheme.

Comunication costs-Member Leave

4000
i i ir s o s
c d
T e moe n

3500
3000
2500
CRT
n c

2000
ECC
1500
m

1000
500
0
0 32 64 128 256 512

Ke y size

Figure 11.Communication Cost
For Member Leave

Consider (Figure.10& 11) there were 256 members in a group our approach consumes only 29%
of Bandwidth when compare to CRTDH and RSADH. So our approach consumes low Bandwidth.
Communica tion cos -Me mbe r join
t

25

20
i i scns
T en eod

RSA
15
DES
CRT
10
m

ECC
5

0
1 2 3 4 5 6
Ke y Siz e

Figure 12. Communication cost for Member Join

For member leave operation also our approach takes less time as the key size for ecc is small
compared to other approaches.

6.4. Computation Costs:
The Computational costs depend on the Number of exponentiations. CRTDH have high
computation costs as it depends on the number of members and group size respectively. The
cost increases as the members and group size increases. But our approach spends a little on this
computation. The number of bits for encryption is very less compared to other keys .Moreover
each user need not store any data key values.



7. CONCLUSION AND FUTURE WORK:
MANETs are much more vulnerable to attack than wired network. This is because of the following
reasons :

Open Medium - Eavesdropping is more easier than in wired network.

Dynamically Changing Network Topology - Mobile Nodes comes and goes from the network,
thereby allowing any malicious node to join the network without being detected.

Cooperative Algorithms - The routing algorithm of MANETs requires mutual trust between nodes
which violates the principles of Network Security.

Lack of Centralized Monitoring - Absence of any centralized infrastructure prohibits any
monitoring agent in the system.

Lack of Clear Line of Defense – There is a lack of clear line of defense - attack prevention may
not suffice.

We have implemented on Multilevel Access Control in a Manet for Defense Messaging System
using Elliptic curve cryptography. Use of Elliptic curve ensures that data is protected and
intruders cannot guess the message. Moreover a single message sent will reach all the classes
which are higher in the hierarchy. The system satisfies the user dynamics and class dynamic. We
have successfully implemented by selecting different elliptic curves. A single elliptic curve can be
used and by changing the generator points and we can perform different encryption. The group
keys are found by the server and forward and backward secrecy is maintained here. The user
level, DataStream level and class level hierarchies are taken care. As future implementation new
methods for improving parameters can be done..

8.REFERENCES
[1] S. Akl and P. Taylor. “Cryptographic solution to a problem of access control in a hierarchy”.
ACM Transactions on Computer Systems, 1(3):239{248,September 1983.

[2] William stallings,” Cryptography and network security Principles and Practices”,Third Edition,
Pearson education.2001

[3] M. Atallah, K. Frikken, and M. Blanton,” Dynamic And Efficient Key Management For Access
Hierarchies”, CERIAS Tech Report 2006-02,Center for Education and Research in,Information
Assurance and Security,Purdue University,

[4] Jason Crampton,”Cryptographically-Enforced Hierarchical Access Control with Multiple Keys”,
Journal of Logic and Algebraic Programming April 1, 2009.

[5] K. Kumar J.Nafeesa Begum, Dr.V.Sumathy, (2009), “A Novel Approach towards cost Effective
Region Based Group Key Agreement Protocol for Ad Hoc Networks” Computational Intelligence,
Communication Systems and Networks ,2009 CICSYN,09 ,July 23-25 2009 published in IEEE
Explore.

[6] K.Kumar, J.Nafeesa Begum ,Dr.V.Sumathy , ,,(2009), “ Efficient Region-Based Group Key
Agreement Protocols for Ad Hoc Networks using Elliptic Curve Cryptography”,IEEE International
Advance Computing Conference(IACC-2009), Thapar University, Patiala March 6-7 . published
in IEEE Explore.



[7] Tim Bauge, White paper on “Ad hoc networking in military scenarios”, Thales Research and
Technology (UK) Limited , May 2004

[8] S. J. MacKinnon, P. D. Taylor, H. Meijer, and S. G. Akl.” An optimal algorithm for assigning
cryptographic keys to control access in a hierarchy”. IEEE Transactions on Computers,
34(9):797.802, Sept. 1985.

[9]S. Chen, Y.-F. Chung, and C.-S. Tian. “A novel key management scheme for dynamic access
control in a user hierarchy”, In COMPSAC, pages 396.397, Sept. 2004.

[10] I. Ray, I. Ray, and N. Narasimhamurthi. A cryptographic solution to implement access control
in a hierarchy and more. In SACMAT ’02: Proceedings of the seventh ACM symposium on
Access control models and technologies, pages 65.73. ACM Press, 2002.

[11] R. S. Sandhu. “Cryptographic implementation of a tree hierarchy for access control”.
Information Processing Letter, 27(2):95.98, Feb. 1988.

[12] G. C. Chick and S. E. Tavares.”Flexible access control with master keys”, Proceedings on
Advances in Cryptology: CRYPTO ’89, LNCS, 435:316.322, 1989.

[13] M. L. Das, A. Saxena, V. P. Gulati, and D. B. Phatak.” Hierarchical key management scheme
using polynomial interpolation”. SIGOPS Operating Systems Review, 39(1):40.47, Jan. 2005.

[14] L. Harn and H. Y. Lin.” A cryptographic key generation scheme for multilevel data security”.
Computers and Security, 9(6):539.546, Oct. 1990.

[15] V. R. L. Shen and T.-S. Chen. ,”A novel key management scheme based on discrete
logarithms and polynomial interpolations”. Computers and Security, 21(2):164.171, Mar. 2002.

[16] M.-S. Hwang, C.-H. Liu, and J.-W. Lo,”. An efficient key assignment for access control in
large partially ordered hierarchy”. Journal of Systems and Software, Feb. 2004.

[17] C. H. Lin. “Dynamic key management scheme for access control in a hierarchy. Computer
Communications,20(15):1381.1385, Dec. 1997.

[18] S. Zhong. A practical key management scheme for access control in a user hierarchy”.
Computers and Security, 21(8):750.759, Nov. 2002.

[19] X. Zou, B. Ramamurthy, and S. Magliveras. “Chinese remainder theorem based hierarchical
access controlfor secure group communications”. Lecture Notes in Computer Science (LNCS),
2229:381.385, Nov. 2001.

[20] X. Zou, B. Ramamurthy, and S. S. Magliveras, editors.” Secure Group Communications over
Data Networks”,Springer, New York, NY, USA, ISBN: 0-387-22970-1, Oct. 2004.


Pankaj U.Lande, Sanjay N. Talbar & G.N. Shinde

FPGA Prototype of Robust Image Watermarking For JPEG
2000 With Dual Detection

Pankaj U.Lande pul@usic.unipune.ernet.in
Dept. Instrumentation Science,
University of Pune,
Pune

Sanjay N. Talbar sntalbar@yahoo.com
S.G.G.S. Institute of Engineering and Technology,
Nanded.

G.N. Shinde shindegn@yahoo.co.in
Indira Gandhi College CIDCO,
Nanded.

Abstract
This paper presents a novel robust invisible watermarking method for still images. The scheme is
implemented on hardware, and it can be incorporated with the lossless JPEG2000 compression
standard. We have implemented Cohen-Daubechies-Favreau (CDF) 5/3 wavelet filters with lifting
scheme which requires less hardware and they are also the basis of lossless JPEG2000. Its
modular structure is well suitable for hardware implementation and it is more efficient use of
power and chip area. The objective of the hardware assisted watermarking is to achieve low
power usage, real-time performance, robust and ease of integration with existing consumer
electronic devices such as scanners, cameras and handy camcorders. The proposed scheme of
watermarking is tested with StirMark software which is a one of the benchmarking software for
watermarking scheme. The experimental result shows that the proposed scheme of watermarking
is robust against most of the geometric attacks such as scaling and rotation. We have proposed a
dual detection technique for watermark detection which is a novelty of our algorithm.

Keywords: CDF 5/3 wavelet, FPGA, watermarking

1. INTRODUCTION
The recent proliferation and success of the internet, together with the availability of relatively
inexpensive digital recording and storage devices has created an environment in which it
becomes very easy to obtain, replicate and distribute digital content without any loss in quality.
This growth of applications in the past decade gave rise to the new set of problems like digital
piracy: illegal copying, use, and distribution of copyrighted digital data. This has become a great
concern to the multimedia content such as music, video and image to the publishing industries,
because technologies or techniques to protect intellectual property rights for digital media and to
prevent unauthorized copying did not exist. Exactly identical copies of digital information, be it
images, text or audio, can be produced and distributed easily. In such a scenario, who is the artist
and who the plagiarist? It’s impossible to tell or was, until now. Digital right management (DRM) is
a collection of technologies and a technique that enables the licensing of digital information
including the multimedia content such as image, video and music.DRM consist of two prominent
technologies those are encryption and watermarking. Encryption technologies can be used to
prevent unauthorized access to digital content. However, encryption has its limitations in
protecting intellectual property rights, because once digital content is decrypted, there is nothing
to prevent an authorized user from illegally replicating it [1][2].



Digital watermarking is a process in which an informed signal (watermark) is incorporated in
multimedia content such as images to protect the owner’s copyright over that content. The
watermark can be later be extracted from a suspected image and be verified in order to identify
the copyright owner. Watermarking technique for paper manufacturing have been in use since the
middle ages, same concept was adopted by digital world and extended this concept for digital
images, video and music.[3]. A watermarking scheme consists of three parts: the watermark, the
encoder, and the decoder. The watermarking algorithm incorporates the watermark in the object,
whereas the verification algorithm authenticates the object by determining the presence of the
watermark and its actual data bits [4].
Watermarking techniques can be divided into various categories in numerous ways [5]. In the
case of still digital images, there are three primary methods for insertion and extraction of a
watermark. These are spatial domain, transform domain and color space methods. The spatial
domain method [6] involves an algorithm that directly operates on the pixel values of the host
image. In the transform domain method the pixel values are transformed into another domain by
applying appropriate transform technique like discrete cosine transform (DCT) [7][9][10],discrete
wavelet transform(DWT)[8][11] and Hadamard transform[12]. A watermark is then embedded by
modifying these coefficients. However it is observed that spatial domain watermarks are weaker
than frequency domain ones [13][14]. A DCT based watermarking algorithm has been described
in many literatures; however DWT based watermarking algorithms are more effective for several
reasons [15].
Wavelet is a small wave whose energy is concentrated in time and still possesses the periodic
characteristics. An arbitrary signal can be analyzed in terms of scaling and translation of a single
mother wavelet function. Properties of wavelet allows both time and frequency analysis of signals
simultaneously. They offer excellent space-frequency localization of salient image features such
as textures and edges. DWT can analyze the data in different scales and resolutions this principal
is called as multi-resolution analysis [16]. DWT decomposed the signal into lower and higher
frequency signal components. The high-frequency content of an image corresponds to a large
coefficient in the detail sub-band. Hence, watermark encoders operating in the wavelet domain
can easily locate the high-frequency features of an image and embed most of the watermark
energy. Such a method of embedding results in an implicit visual masking of the watermark,
because the ability of human visual system (HVS) to detect high frequency signals is limited [17].
It is also a basis of a compression standards like JPEG2000 [18] and MPEG-4[19].

1.1 Related Work
Software approach for image watermarking have been proposed in many literatures; but
hardware implementations has few advantages over software approach such as
1. It gives optimized specific design which is a small, fast, and potentially cheap
watermarking unit.

2. It is most suitable for real-time applications, where the computation time is
deterministic and short.

3. Hardware based watermarking unit can be easily integrated with digital cameras and
scanners, graphics processing units etc.

4. Hardware watermarking unit consumes lesser power than software, which requires a
general purpose processor so that they are ideal for battery operated applications.

5. The cost is low compared to that software used explicitly for watermarking; this is
because a hardware based watermarking unit can be monolithically built on a single
unified system in the context of system-on-chip (SoC) technology.



6. The hardware can be implemented as a soft core expressed in the structural
hardware description language like VHDL and Verilog. The soft core can be modified
as algorithm changes and can be resynthesized into new silicon technology.

A hardware based watermarking is presented in few literatures illustrated bellow.
Seo and Kim [20] presented a field programmable gate array (FPGA) based implementation of
blind and invisible watermarking on Altera FPGA. This watermarking algorithm was presented in
DCT domain and the DC coefficients are replaced by watermark. The two dimensional DCT was
calculated for one or more than one bit planes and the DC coefficients are replaced in such a way
that it will be imperceptible to human eyes. The watermarking algorithm was integrated with
JPEG2000 encoder and it operates on 66MHz.
An FPGA based invisible robust spatial domain watermarking is described in [21].the watermark
insertion is carried out by replacing original image pixel value by watermark encoding function.
The watermark is generated through user key. The watermarking scheme is evaluated by
standard benchmark like StirMark software. The original image is required for watermark
detection. The algorithm was implemented on XCV50-BG256-6 device from Xilinx and operated
on 50.398MHz.
An FPGA prototype of Biometric based watermarking is described in [22]. The algorithm work for
both gray and color image and the biometric image is selected as watermark. The original image
is divided in 8×8 blocks and DCT is calculated for each block. The biometric image (watermark) is
divided into blocks and embedded into perceptually significant region of cover image. This
approach makes the watermark robust against the common signal processing attacks. Original
image is required watermark detection. The prototype was modeled using VHDL and
implemented on XC2V500-6FG256 device from Xilinx.
Saraju P. Mohanty et. al. [23] proposed a novel algorithm for encrypted watermarking based on
block-wise DCT. The watermarking can work for gray scale image as well as color image. In the
case of color image the cover image in RGB format is converted into YCbCr and the Y
component is selected for watermarking. The image is divided in 8×8 blocks and DCT is
calculated for each block. The encrypted watermark is embedded into transformed image by four
different embedding factors. The embedding strength factor is chosen such that the image quality
will not degrade. The watermark detection process requires original image. The block-wise DCT
is computed for both image and the difference is calculated to detect a watermark. The extracted
watermark is compared with original watermark to authenticate the suspected image.
Image adaptive watermarking and its hardware architecture is described in [12]. The proposed
scheme of watermarking is invisible and robust against JPEG attacks. Cover image is divided in
8×8 blocks and DHT is calculated. PN sequence is generated through user key and embedded
into DHT coefficients. The strength factor is calculated from quantization table for DHT domain.
Watermark detection method is blind. The proposed method is robust against the common signal
processing attacks like median filtering and noise addition. The algorithm was implemented on
XC3SD1800A-4FGG676C and functional simulation was performed using Xilinx tools. The chip
was tested using hardware co-simulation which was run at 33.3MHz.
In this paper we have presented a watermarking scheme using the CDF 5/3 wavelet filter which
can be incorporated with JPEG2000 lossless image compression. Hardware architecture was
implemented on FPGA. The proposed scheme is an invisible robust wavelet domain
watermarking method. We have also proposed a dual watermark detection technique that is the
watermark can be detected by blind and non blind method to verify the suspected image. The
blind watermark detection can be used with the images from digital cameras where the original
image is not present. The non-blind technique can be used with the digital scanners where the
original image is present.
The scheme described in [24] is used to implement CDF 5/3 wavelet filters. The proposed
architecture uses the lifting scheme technique and provides advantages that include small
memory requirements, fixed-point arithmetic implementation, and a small number of arithmetic
computations. The chip was modeled using Verilog and a function simulation was performed.
This chip was tested using AccelDSP in hardware in the loop (HIL) arrangement. The proposed
scheme is robust against several geometric attacks. We have tested our watermarking scheme
using standard benchmark such as StirMark software.



2. PROPOSED WATERMARKING SCHEME
The proposed scheme is based on CDF 5/3 wavelet filters which is the basis of lossless
JPEG2000 compression standard. The new still compression image standard, JPEG2000 has
emerged with a number of significant features that would allow it to be used efficiently over a wide
variety of images. The JPEG2000 standard exhibits a lot of features, the most significant being
the possibility to define regions of interest in an image, the spatial and SNR scalability, the error
resilience and the possibility of intellectual property rights protection. Interestingly enough, all
these features are incorporated within a unified algorithm. This compression standard uses the
Cohen-Daubechies-Favreau (CDF) 5/3 and CDF 9/7 DWT for lossless and lossy image
compression respectively. Since JPEG2000 is the newest version of one of the most popular
image formats and it includes the DWT, efficient VLSI implementations of DWT processors
became more and more important.
We have used the lifting scheme descried in[24].the advantages of using lifting scheme is that,
the number of multiplications and additions compared to the filter-bank implementation are
reduced resulting in more efficient use of power and chip area. The modular structure is well
suitable for hardware implementation. The lifting scheme calculates the DWT using spatial
domain analysis, and consists of a series of Split, Predict and Update steps. The split step
separates odd and even samples, and predict step predicts values in the odd set where α= -0.5
as the predict step coefficient. The Update step uses the new wavelet coefficients in the odd set
to update the even set, where β=0.25 as the update step coefficient. Lifting scheme is shown in
figure 1 and Lifting operation for the CDF 5/3 synthesis filter is shown in Figure 1.

FIGURE 1: CDF 5/3 with lifting scheme

The watermarking algorithm embeds same multiple watermarks in cover image. The cover image
I is divided into non-overlapping blocks of size B×B. CDF 5/3 wavelet transform is calculated for
block separately. A binary watermark is embedded into cover image using equation (1).

IW,N (x, y)= IN (x, y)+a ×W(x, y) (1)

Where ‘a’ is gain

IW,N is a Nth block of watermarked image and W is a binary watermark logo. x and y are index
numbers.

Cover CDF 5/3 CDF 5/3
∑ Inverse Wavelet Transform
Image Wavelet Transform

Watermark × Watermarked
Image

Gain

FIGURE 2: Watermarking Scheme

2.1 Hardware Architecture for Proposed Scheme



In this section, we discuss the hardware architecture for the scheme discussed in the previous
section. The watermarking chip mainly consists of a block processing unit and control.

2.1.2 Block processing unit
The block processing unit considers the original image block as input. Image block is wavelet
transformed and the watermark is embedded using equation (1). This unit consists of CDF5/3
wavelet filters and watermarking unit. To meet the real time constrain, we have used two filters in
parallel to calculate forward and inverse transform. In order to calculate the 2D wavelet, these
filters first calculate the coefficients first row-wise and then column-wise. The intermediate results
are stored in the memory. Inverse wavelet is calculated in similar manner.
2.1.3 Watermarking unit
The watermarking unit consists of a multiplier and adder. The watermark is embedded using
equation (1). Because a multiplier requires more hardware, only one multiplier is implemented.
The wavelet transformed block is fed serially to the watermarking unit. The gain is multiplied by
watermark and added to the wavelet transformed coefficients. The intermediate results are stored
in the memory.
2.1.4 Control unit
The control unit generates the necessary control signals for the entire system during the
watermarking process. The control unit generates four main signals and these signals are as
follows
CLOCK_ENABLE

S1
OUTPUT_AVIL INPUT_AVIL

S7 S2

S6 S3

S5 S4

FIGURE 3: Control Unit as FSM

INPUT_AVIL: Image block is available at input.
OUTPUT_AVIL: Watermarked image block is available at output
CLOCK: Clock signal for chip
CLOCK_ENABLE: When clock enable is high chip is in an active mode for processing
This unit undergoes seven states in each state; the particular task is performed in each state and
the finite state machine (FSM) begins to the next state. Figure (3) shows the state diagram of
FSM.
S1: if the clock enable is high and INPUT_AVIL is high then read image block
S2: calculate DWT
S3: read the watermark
S4: multiply the watermark with ‘a’
S5: embedded the watermark
S6: calculate inverse DWT
S7: generate OUTPUT_AVIL signal



2.1.5 Watermark detection

The watermark detection algorithm is implemented using MatLab. The watermark can be
detected using two methods blind and non-blind. In non blind method original and watermarked
image both are required to detect a watermark. The suspected image and original image are
divided into B×B blocks, and DWT coefficients are calculated for both images. The watermark is
recovered using equation (4).

1 if IW,N (x,y) - I N (x, y)  
W(i, j)   (4)
0 other wise

 represents threshold for blind detection
In the blind watermark detection method the binary logo image is considered as PN
sequence and the correlation between the suspected image and watermark is calculated. The
suspected image is divided into B×B blocks, and DWT coefficients are calculated. The correlation
between encrypted watermark and wavelet transformed block is calculated using equation (5)

(5)

If > ρ then the watermark is detected. ρ is threshold for blind detection.

If > ρ then the watermark is detected. ρ is threshold for blind detection.

3. EXPERIMENTAL RESULTS

3.1 Synthesis and Implementation

The chip was modeled using a Verilog and functional simulation was performed. The code was
synthesized on Xilinx Spartan-3A technology on XC3SD1800A-4FGG676C device using the
AccelDSP. The results are verified by hardware in the loop (HIL) configuration using AccelDSP.
The HIL was run at 33.3 MHz clock frequency, and the samples were fed to the target device at a
rate of 319.585 Ksps through a JTAG USB cable. The design utilizes 2 startup clock cycles and
single clock cycles per function call. The device utilization summary is given in Table 1.

Logic Utilization Used Available Utilization %
Number of Slices 628 16640 3.77 %
Number of Slice Flip Flops 290 33280 0.87 %
Number of 4 input LUTs 1077 33280 3.23 %
Number of bonded IOBs 293 309 94.82 %
Number of GCLKs 1 24 4.16 %

TABLE 1: Device Utilization Summary



FIGURE 4: RTL of Watermarking Chip

3.2 Image Quality Measures

In [25] Kutter and Petitcolas have discussed various parameters to estimate any watermarking
scheme. For fair benchmarking and performance evaluation, the visual degradation due to
embedding is an important issue. Most distortion measure (quality metrics) used in visual
information processing belongs to a group of difference distortion measures. The watermark
images are acceptable to the human visual system if the distortion introduced due to
watermarking is less.
The various performance evaluations metrics such as PSNR (db), Image Fidelity (IF), Normalized
cross correlation, correlation quality etc. are calculated. Results for few popular images are given
in Table 2.

Quality Measures Lena Mandril Woman
Mean square error 6.80 6.74 6.80
PSNR 39.79 39.84 39.80
Normalized cross 1 1 1
correlation
Average Difference -0.8135 -0.8055 -0.8146
Structural content 0.98 0.99 0.98
Maximum difference 3 3 3
Normalized absolute 0.031 0.017 0.017
error
Image Fidelity 1 1 1
correlation quality 1 1 1

TABLE 2. Image Quality Measures



3.3 Performance Evaluations on Various Attacks

In this section, we evaluate the performance of the watermarking algorithm against various
attacks using standard benchmark software. The StirMark software includes several attacks such
as JPEG compression geometric transformation, noise addition etc. The geometric attacks
includes rotation, cropping, scaling and geometric transformation with medium compression.
Some of the results of these evaluations for blind and non-blind detection are summarized in
Table 3. For blind detection threshold is ρ=0.85. These results indicate that the proposed
watermarking scheme is robust against the geometric attacks.

The proposed scheme of watermarking embeds multiple watermarks in cover image. The
objective was, at least a single watermark will survive after attacks. In detection algorithm all the
watermarks are detected, and the watermark which is having highest correlation with the original
watermark is treated as the recovered watermark. We have also proposed a dual watermark
detection technique. The watermark can be detected by blind or non-blind method and both
detection techniques can be used to verify suspected image. Scheme implements several
watermark in the cover image, due to which scheme is robust against various geometric attacks.

4. Conclusion
In this paper, we proposed a novel invisible image watermarking algorithm and developed
efficient the hardware architecture which can be used with JPEG2000. The watermarking scheme
utilizes minimal hardware resources as it can be seen from the device utilization summary table.
Because of the lifting scheme is used in CDF 5/3 filters it requires minimum hardware and it
requires less clock cycles. The experimental results showed that the proposed scheme of
watermarking scheme is imperceptible and robust against geometric attacks. The proposed
algorithm outperforms than the presented algorithm in [12][21]. This was achieved because of
space and frequency localizing property that is the characteristics of the discrete wavelet
transform. In the feature we want to develop a image adaptive watermarking hardware using
fuzzy logic or neural network.

Attacks Non – Blind Detection
Blind Detection

AFFINE_2 Max corr value =0.86

CROP_25 Max corr value =0.87



MEDIAN_3 Max corr value 0.88

RML_90 Max corr value =1

ROTCROP_-75 Max corr value 1

ROTSCALE_0.5 Max corr value 1

TABLE 3. Performance Evaluation

5. REFERENCES

1. Er-Hsinen, “Literature Survey on Digital Image Watermarking,”EE381K-Multidimentional
Signal Processing 8/19/98.

2. S. Katzenbeisser and F. A. P. Petitcolas: Information Hiding techniques for
steganography and digital watermarking, Artech House, Inc., MA, USA, 2000.

3. N. Memon and P. W. Wong.:Protecting Digital Media Content. Communications of the
ACM, vol. 41, no. 7, pp. 34–43, Jul 1998.

4. C.C. Chang and J. C. Chuan, “An image intellectual property protection scheme for gray-
level images using visual secret sharing strategy,” Pattern Reconition Letters, vol. 23, pp.
931-941, June 2002.

5. S. P. Mohanty.:Watermarking of Digital Images.M.S. Thesis, Indian Institute of Science,
Bangalore, India, 1999.
6. N. Nikolaidis, I. Pitas, “Robust Image Watermarking in Spatial Domain”, International
journal of signal processing, 66(3),385-403,1998.



7. Pankaj U. Lande,Sanjay N.Talbar and G.N. shinde “Adaptive DCT Domain Watermarking
For Still Images”, Internatational Conference RACE-07,Bikaner, Rajastan, India

8. Pankaj U.lande,Sanjay N.Talbar and G.N. shinde “Hiding A Digital Watermark Using
Spread Spectrum At Multi-Resolution Representation”, Internatational conference
ACVIT07, Aurangabad, India.

9. Juan R.Hernandez,Martin Amado,Fernando Perez-Gonzalez ”DCT Domain watermarking
technique for still Image :Detectors Performance analysis and New Structure”,IEEE
transaction on image processing,VOL.9,no.1,Jan 2000.

10. Juan R. Hernandez, Martin Amado, Fernando Perez-Gonzalez “DCT Domain
watermarking technique for still Image: Detectors Performance analysis and New
Structure”, IEEE Transaction on Image Processing, VOL.9,No.1, Jan 2000.

11. Pik Wah Chan,Michael R.lyu and Roland T. Chin, “A Novel scheme For Hybrid Digital
Video Watermarking : Approach, Evaluation And Experimentation”, IEEE Transactions on
circuits and system for video technology,VOL 15,No. 12,Dec 2005.

12. Pankaj U.Lande,S.N. Talbar,G.N. Shinde, “FPGA iplimentation of image adaptive
watermarking using human visual model”,ICGST-PDCS,Vol.9,Issue1,Oct. 2009.

13. I. J. Cox, J. Kilian, T. Shamoon, T. Leighton, Secure Spread Spectrum Watermarking of
Images, Audio and Video, in: Proc IEEE International Conf on Image Processing, Vol. 3,
1996, pp. 243–246.

14. I. J. Cox, J. Kilian, T. Shamoon, T. Leighton, A Secure Robust Watermarking for
Multimedia, in: Proc. of First International Workshop on Information Hiding, Vol. 1174,
1996, pp. 185–206.

15. P. Meerwald and A. Uhl,(2001) “A survey of wavelet-domain watermarking algorithms,”
Proceedings of SPIE, Electronic Imaging, Security and Watermarking of Multimedia
Contents III, San Jose, California, vol. 4314.

16. Burrus, C., Gopinath, R., and Guo, H.: Introduction to Wavelets and Wavelet Transforms:
A Primer.: Prentice Hall 1998.

17. R. Dugad, K. Ratakonda, and N. Ahuja,(1998) “A new wavelet-based scheme for
watermarking images,” Proceedings of the IEEE International Conference on Image
Processing, ICIP ’98, Chicago, Illinois, pp. 419-423.

18. D. Taubman and M. Marcellin.: JPEG2000: Image compression fundamentals, standards,
and practice.:springer, 2002.

19. 19.T. Ebrahimi and F. Pereira.: The MPEG-4 Book.: Prentice Hall 2002.

20. 20 Y. H. Seo, D. W. Kim, Real-Time Blind Watermarking Algorithm and its Hardware
Implementation for Motion JPEG2000 Image Codec, in: Proceedings of the 1stWorkshop
on Embedded Systems for Real-Time Multimedia, 2003, pp. 88–93.

21. S. P. Mohanty, R. K. C., S. Nayak, FPGA Based Implementation of an Invisible-Robust
Image WatermarkingEncoder, in: Lecture Notes in Computer Science, Vol. 3356, 2004,
pp. 344–353.



22. S. P. Mohanty, O. B. Adamo, and E. Kougianos, "VLSI Architecture of an Invisible
Watermarking Unit for a Biometric-Based Security System in a Digital Camera", in
Proceedings of the 25th IEEE International Conference on Consumer Electronics (ICCE),
pp. 485-486, 2007.

23. S. P. Mohanty, “A Secure Digital Camera Architecture for Integrated Real-Time Digital
Rights Management”, Elsevier Journal of Systems Architecture (JSA), Volume 55, Issues
10-12, October-December 2009, pp. 468-480.Kanchan H. Wagh, Pravin K. Dakhole,
Vinod G. Adhau.: Design & Implementation of JPEG2000 Encoder using VHDL.
Proceedings of the World Congress on Engineering 2008 Vol I, WCE 2008, , London,
U.K July 2 - 4, 2008.

24. Kanchan H. Wagh, Pravin K. Dakhole, Vinod G. Adhau.: Design & Implementation of
JPEG2000 Encoder using VHDL. Proceedings of the World Congress on Engineering
2008 Vol I, WCE 2008, , London, U.K July 2 - 4, 2008.
25. M. Kutter and F.A.
Petitcolas, “A Fair Benchmark for Image Watermarking Systems”,Electronic imaging
,Security and Watermarking of Multimedia Contents, VOL. 3657, 25-32,1999.


S. Gkarafli & A.A. Economides

Comparing the Proof by Knowledge Authentication Techniques

Stamati Gkarafli gkaraflistamati@yahoo.gr
Information Systems Department
University of Macedonia
Thessaloniki 54006, GREECE

Anastasios A. Economides economid@uom.gr
Information Systems Department
University of Macedonia
Thessaloniki 54006, GREECE

Abstract

This paper presents a survey of proof by knowledge authentication techniques (text
passwords, visual passwords and graphical passwords). Both new methods are
more memorable, as people have to remember images and not characters and
graphical passwords are also more secure. A total of 100 users participated in our
survey, who after getting informed about the new authentication methods, they
answered the questions of our questionnaire. Based on their answers, all
participants have many passwords for their everyday needs and they try to select
passwords that are not only memorable, but also secure. Unfortunately, they can not
deal with proper password selection and they become victims of dictionary attacks.
Understanding this situation, participants were very positive in learning more about
the new authentication methods. They found both techniques memorable and
friendly – visual passwords at most. However, they found graphical passwords a bit
more complex and difficult to learn how to use them, something that they can
overcome with small practice.

Keywords: graphical passwords; text passwords; user authentication methods;
visual passwords.

1. INTRODUCTION
Access control includes the user’s identification and authentication, authorization, audit and
accountability. Various access control models have been proposed in the past [e.g. 1-5].

Authentication is the process of confirming or not the user’s identity. Jansen [6] distinguished the
authentication techniques into the following three categories:
Proof by knowledge techniques, which are based on specific information that an individual has (e.g.
PIN- personal identification number, text-passwords).
Proof by property techniques, which are based on a certain property that the user has (e.g.
biometrics, fingerprint verification, voice verification, iris scanning) [e.g. 7].
Proof by possession techniques, which are based on the possession of an object that an individual
has (e.g. smart cards, digital certificates, security token).

Text passwords represent the authentication method that is mainly used by all users today.
Nevertheless, most times users select passwords that are memorable and as a result easy to be
cracked [8]. This problem is very serious if one looks at some case studies that were conducted.
According to a security team in a large company, they managed to crack 80% of the passwords [9].
Also, based on Klein’s case study [10], 25% of 14.000 passwords were cracked using a small
dictionary of 3 million words. According to these results one can assume that even if these methods



are very popular, they can cause serious problems to the users. Some of the usual problems are the
following:
 Users choose passwords that are very short in length.
 Users choose passwords that are easy to remember.
 Users write passwords down or share them with others, in order to remember them easier.
 Users use the same passwords for different applications.

Text passwords are very vulnerable to “dictionary attacks” (automated attacks using tools that can
crack the passwords that are common words, names or dates).

2. VISUAL AND GRAPHICAL PASSWORDS
Considering all those problems of text passwords, researchers invented other proof by knowledge
authentication methods: visual and graphical passwords. These methods have many advantages that
are described below [11, 12, 13]:
A sequence of pictures is more memorable than a sequence of characters.
Pictures are independent from user’s language.
There do not exist yet special dictionaries for a dictionary attack and it is very difficult to be
constructed (especially for graphical passwords that have a very large password space).
Automated attacks are difficult to take place.

Except from all these advantages Renaud and De Angeli [14] referred the shoulder surfing problem
as the main disadvantage of the new methods. This happens when someone is looking over
someone’s shoulders, during the login process [15, 16].

Next, we describe the two new authentication methods and the applications that have been created
for each one of them.

2.1 Visual Passwords
Visual passwords are passwords that are created by selecting a sequence of images [5, 17]. Based
on this idea, there was developed a number of commercial products that we will refer and analyze in
the text below.

Passfaces
Real User Corporation developed a product named “Passfaces” [14, 18]. The basic idea is that the
user must select four images of human faces, in a specific order, from a database. This sequence of
images will be his secret password.

During the login process, the user sees a grid of nine faces, but only one of those faces belongs to
the password that he made. He clicks on it, another set of nine images appears and he must again
recognize and click on the image that is included in his password. This procedure is repeated until the
user clicks on all four pictures of his password. If he clicks on all four pictures correctly, the system
allows him to enter.

The whole idea is based on the fact that people recall much easier human pictures than any other
kind. This is especially true if each user has the opportunity to construct his own database with his
personal pictures which are familiar to him and much harder to be guessed by an attacker.

FIGURE 1: Passfaces



Story Scheme
Story Scheme is an application similar to Passfaces. The only difference is that the images would be
not only human faces, but also everyday objects, animals, food, sports, cars, or even people. In this
case, the user has to think a story with the pictures that are in his database. Having his story in his
mind, he selects a sequence of images [19].

FIGURE 2: Story Scheme

Déjà Vu
Dhamija and Perrig developed “Déjà Vu” for user authentication [20]. At the beginning the user
creates his “image portfolio”, by selecting a set of p images out of a much bigger set.

During the login process, the Déjà Vu presents to the user a set of n images consisting of:
x images that belong to the users’ image portfolio, and
y other images, that we call “decoy images”.

At this point the user has to make a click on all images that he recognizes as belonging into his image
portfolio. If he succeeds in it, the user will enter the system successfully.

FIGURE 3: Déjà Vu

Something that is very important in Déjà Vu scheme is the type of images that are used. The images
are based on Andrej Bauer’s Random Art [21], which can generate random abstract images. When
the user attempts to make his image portfolio, the system gives to this specific process an initial seed.
With this seed, Random Art generates a random mathematical formula, which defines the color of
each pixel in an image. Therefore, each user has his personal random pictures that were created from
the initial seed.

So, it is obvious that this type of images makes the whole system much more secure, since it is very
difficult for someone who observes the login process to remember them. Moreover, the system does
not have to store each image pixel-by-pixel, but it stores only the initial seed.

Picture Password



Picture Password authentication mechanism is another application for visual login, which was
developed by Jansen [5, 22].

The images are grouped into different categories according to the theme that they represent. Theme
examples include Cats & Dogs, Sea, Landscapes, Sports, Faces, Transportation Means, etc. In order
to create a password, the user has to choose one of these themes and afterwards a sequence of
images from this theme.

Each image corresponds to an element of an alphabet. However, the user does not have to
remember a sequence of random characters, but a sequence of images, something that is much
easier to recall.

There are two different ways to choose the sequence of images:
Individual Selection: each image represents one element in the alphabet
Paired Selection: two images are combined (more often with drag and drop) and their coupling
represents one element in the alphabet.

FIGURE 4: Picture Password.

Passlogix – Passpoints
Passlogix [23] is based on Blonder’s idea [24] who proposed that during authentication, the user must
click on several locations in an image. In the Passlogix implementation, the user must click on a
sequence of items in the image he sees on his screen in order to create his password. To make a
successful login to the system, he has to click on the same items in the correct order. A problem that
we face here is how we know if an item is clicked, if we click on its edge. For this reason, there were
defined invisible boundaries which indicate whether an item is clicked by the mouse or not.

FIGURE 5: Passlogix.

An extension of this idea is the “Passpoint” system developed by Wiedenbeck et al. [25, 26, 27]. In
this implementation, the user is able to make a click at any place on an image, and not only on a
certain object. To achieve this they eliminated the boundaries between objects and put a tolerance
around each pixel that is selected. As a result, the only thing that a user must do to make a login and
enter the system is to click within the tolerance of the pixels that he chooses, in the correct order.



It is important to be mentioned here that complex pictures can have hundreds of memorable points.
So, with 5 or 6 clicks, a user can make more passwords than that used today with 8 characters. So,
the possible password space with this method is very large, and the whole process becomes very
safe.

FIGURE 6: Passpoints.

2.2 Solving the Shoulder Surfing problem
As we have already mentioned, one main problem that visual passwords have is the “shoulder surfing
problem”. That is watching over someone’s shoulders as he tries to login a system [16]. Next, we
present three methods developed by Sobrado and Birget [15] in order to solve this problem.

Triangle Scheme
According to this scheme, the user sees on his screen a set of N objects in a random position each
time. He creates his password by selecting K pass-objects that will consist from now on, the user’s
portfolio.

During the login process the user is able to see on his screen a set of L images with L<N. To login
correctly he must recognize the 3 pass-objects that are depicted and make a click inside the invisible
triangle that is created.

If this process is done only once, then it would be very easy for someone to click inside the triangle by
chance. So, the process is repeated for 8 – 10 times for each login, with the purpose to reduce the
probability of randomly clicking on the correct region.

FIGURE 7: Triangle Scheme.

Movable Frame Scheme
Based on the same ideas with the previous scheme, in the “Movable Frame Scheme”, the user has to
recognize 3 of the pass objects that he has already chose. This time, his job is to move the frame
around (Figure 8) until the pass object on the frame lines up with the other two.

Of course, the process is again repeated for several times, as it was also mentioned above, in order
to avoid lining up the correct items by chance.



FIGURE 8: Movable Frame Scheme.

Other Geometric Configurations
Similarly, the third method follows the same rules. However, in this case the user has to recognize 4
pass objects and make a click at the intersection of the virtual lines that are formed by connecting
these objects.

Of course, it is obvious, that there is a tolerance of the pixel that the user clicks, in order not to have
arguments about the exact point that he chooses each time.

FIGURE 9: Other Geometric configurations.

2.3 Graphical Passwords
In this section, we analyze graphical passwords. Here, the user has to draw a personal design that
will be from now on his secret password. More precisely, we examine the DAS scheme and its
extension Multi-grid Passwords.

Draw-a-Secret (DAS) Scheme
Jermyn et al. [22] proposed a scheme, called DAS (Draw-a-Secret), which allows the user to draw a
simple design on a grid. On his screen the user is able to see a rectangular grid of size G  G. Then,
he has to draw a sequence of lines in this grid. This drawing will represent his password.

For example, let consider a grid of size 3  3 (Figure 10). In order to create his password, the user
has to draw a sequence of lines. Then, at the login process, he must draw the same lines again, in
the same order. For this reason the drawing is mapped to a sequence of coordinate pairs by making a
list of the cells through which the drawing passes in the correct order, separated by pen-up events,
when the user raises his pen and continues from another point. In Figure 10, we have the following
sequence that we made by the drawing:
(1,2), (1,1), (2,1), (2,2), (3,2), pen-up, (2,3), (1,3).

According to this, we must underline that what counts for a user in order to make a successful login is
not the exact draw that he has made, but the sequence of cells from which his pen passes in
combination with the pen-up events.



FIGURE 10: DAS Scheme.
In Figure 11 we can see another example of the DAS Scheme, its internal representation, the bit
string that is created and an unsuccessful login because of a shift error that was made.

An important point is that a one-way hash function is applied to the bit string that is created and the
result is stored to the server of the system. So, when the user tries to login the system, he draws his
password, then the hash function is applied to it, and the result is compared with the stored result. If
these two are the same, then the login is successful, otherwise it is not.

As we can understand this process makes the whole scheme very safe. The system does not know
each user’s password and as the hash function is one-way we cannot compute the initial design from
the result that is stored.

FIGURE 11: DAS Scheme. Internal representation of our draw and a failed login
because of a shift error [28].

a) Multi-grid Passwords
Nali and Thorpe [29] performed a survey to investigate the reasons that could make DAS passwords
vulnerable to attacks. According to them, this idea could be true if we consider that many users draw
passwords with predictable characteristics, which means that these are centered or symmetric. As a
result symmetrical or centered DAS passwords can reduce very much the password space and help
attackers in creating dictionaries [30, 31]. Based on this remark, Birget et al. [30] referred problems
with the DAS scheme because of uncertainty in the clicking regions.

Chalkias et al. [33] as well as Alexiadis et al. [34] made an extension to the DAS scheme with the aim
to reduce these facts that have very serious implications to the security of the system. They proposed
multi-grid passwords by using nested grids in the initial one (Figure 12). The main reasons that users
fail to confirm their passwords are that they forget their stroke order and they mark adjacent cells
instead of the correct ones.

With multi-grid passwords the users are able to create more complicated passwords that at the same
time are more memorable. For example, using grids like in Figure 12, the user has more points of
focus to do his drawing, and so he does not need to make his design in the center of the grid or
symmetrically. With multi-grid passwords the users are able to create more complicated passwords
that at the same time are more memorable. This can be explained if we examine Figure 12. With grids



like those, we have more points of focus to do our drawing, and so the user does not need to make
his design in the center of the grid or symmetrically.

FIGURE 12: Multi-grid passwords [33].

3. RELATED WORK
As we have already mentioned, text passwords nowadays are very unsafe and can cause serious
security problems to the users. Based on this fact and taking into account the new authentication
methods that were invented, many researchers have worked on them, with the aim to compare them
and find what effect they may have in our society.

According to Irakleous et al. [35], 59% of the participants have 2 -5 passwords, while 65% of them
use at least one of them daily. Also, Tribelhorn [36] showed that all users are concerning very much
for the security of their passwords. For this reason, many users try to create passwords that are
difficult to be cracked, by selecting 8 or more characters. But just this effort is not enough, because
users want their passwords to be also memorable. So they create passwords that are consisted of
ordinary words or dates, making them really vulnerable to dictionary attacks.

All these lead the researchers to the new authentication methods, visual and graphical passwords.
Many applications were created based first of all on the image selection of visual passwords. Kim and
Kwon [37] found that having in visual passwords images with known faces, make them more
memorable that having landscapes or random faces. Davis et al. [19] examined the images that males
and females choose and concluded that females chose animals or food and males choose women
and sports. Dhamija and Perrig [20] working on Déjà vu scheme, found out that after some practice,
users have very good results in remembering the passwords with simple photos, but also the
passwords that are based on random art technique.

Really opposed to these results, Jansen [22] showed that visual passwords are not safer than the
text, as users tend to select a small number of images (usually 4 or 5), creating this way passwords
easy to be cracked. Wiedenbeck et al. [25, 26, 27] showed that graphical passwords may be more
time-consuming or more difficult to be confirmed successfully, but that after some practice the
situation changes and above all the passwords that are created are really safe.

Further examining the graphical passwords, Goldberg et al. [38] found out that with practice users can
have the best results. They make more successful logins, something that is confirmed by the fact that
72% of them agreed that this method is easier to remember than the text passwords.

Nali and Thrope [29] examined the drawbacks of graphical passwords which include that users draw
designs that are centred or symmetrical, something that decreases their security. As a solution to this
situation, Chalkias et al. [33] proposed a multi-grid password scheme and they made a comparison for
centred and symmetrical drawings, between non-technical and technical users. According to their
survey, using a multi-grid scheme fewer participants created centred and symmetrical passwords and
as a result, more users were able to create safer passwords. Weiss and Del Luca [39] proposed
PassShapes. In this system users authenticate themselves to a computing system by drawing simple
geometric shapes constructed of an arbitrary combination of eight different strokes. Also, Eljetlawi and
Ithnin [40] designed Jetafida focusing on the usability features of this graphical password system.
Everitt et al. [41] found that the frequency of access to a graphical password, interference resulting



from interleaving access to multiple graphical passwords, and patterns of access while training
multiple graphical passwords significantly impact the ease of authenticating using multiple facial
graphical passwords. Chiasson et al. [42] compared the recall of multiple text passwords with recall of
multiple click-based graphical passwords. In a one-hour session (short-term), they found that
participants in the graphical password condition coped significantly better than those in the text
password condition. Similarly, Ozok and Holden [43] compared alphanumeric and graphical
passwords. Johnson and Werner [44] found that passcodes are more memorable than alphanumeric
passwords over extended retention intervals. Finally, various classification systems of graphical
passwords were proposed [45, 46].

In our survey, in opposition to those that we have already mentioned, we made a comparison
between the three “proof by knowledge” authentication techniques, text, visual and graphical
passwords and not just between two or three specific applications that are listed either in the visual or
graphical method. Also 100 users were participated with the aim to obtain results that are significant
statistically, while in the other surveys there were participated up to 40 users. Finally, to have more
general results, the users that participated were from 18 to 60 years old (not only students in colleges
and universities), in order to find out the impression and the effect that the new methods have in the
whole society and not just in a part of it.

Furthermore, we created a comprehensive study, to examine many issues that refer to visual and
graphical passwords, comparing them with the traditional authentication method that is text
passwords.

4. METHODOLOGY
4.1 Questionnaire
To conduct our survey, we created a questionnaire to obtain users’ opinion. Our questionnaire is a
completely new questionnaire that is referred to characteristics that text, visual and graphical
passwords have. More precisely, the information that we collected with the questionnaire are the
following:
how important are passwords for the users
what problems do text passwords have
are they positive in using visual and graphical passwords
user’s personality according to their personal opinion
users’ opinion about text, visual and graphical passwords in terms of memorization, difficulty in
learning their use and friendliness.

Analyzing users’ answers to the questionnaire, we discovered their opinion about the traditional text
passwords and the new visual and graphical passwords. In addition, we examined at what degree our
society is able to change its habits and accept without any problems, the new authentication methods
that are really safer.

4.2 Participants
In our study 100 people were participated. They already knew and used PINs and text passwords.
There were included 49 men and 51 women, from 18 to 60 years old. More precisely, if we divide the
users into three categories according to their age, we have: 59 users from 18 to 30 years old, 20
users up to 45 years old and 21 users over 46 years old and until 60. The younger participants were
students or post graduated students in the University of Macedonia in Thessaloniki, in the Aristotle
University in Thessaloniki and in the ΤΕΙ of Larissa. The elder ones were teachers in schools and
Universities, people that work in public services, in the private sector, in manufactures or in factories,
in Thessaloniki and in Larissa. All these users were found in their work, something that made really
difficult the whole process of collecting their answers. This is because we had to make the procedure
that we will describe analytically below, separately to 3 - 4 users, as it was infeasible to find more
people free at the same time, to deal with our questionnaire.

What was very important to us was that from the beginning of our survey every participant was very
positive in helping us to carry it out. Especially after listening to a small introduction with information
about the new methods and the danger that traditional text passwords hide, they were very willing to
participate and in this way help us overcome all these problems.



4.3 Procedures
As we have already mentioned, in order to accomplish our survey, we have created a questionnaire
that referred to text, visual and graphical passwords. The whole procedure of our survey is divided
into two different stages. At the first stage, we gave a short seminar (as many of them were in their
work) to the participants that included an introduction to the text passwords and their problems and an
analysis of the new authentication methods, visual and graphical passwords.

After understanding all these points, all the participants were asked to fill in the questionnaire that we
have prepared, with the aim to find out their personal opinion about the three authentication methods.
Finally, after collecting all the questionnaires, we analyzed the users’ answers and we present the
results in the text below.

5. RESULTS
In this section, we present the results of the survey regarding text, visual and graphical passwords.
We present the results in six categories: 1) password importance, 2) problems with text passwords, 3)
the new authentication techniques, 4) comparison among the three methods, 5) making safer
graphical passwords.

5.1 Are passwords important to our life?
First of all, we were interested in identifying the importance of passwords from the users’ point of
view. We asked participants about the number of passwords they use in their everyday activities. 47%
of them have one to four passwords, 34% have 5 to 7 passwords, and 19% have more than 8
passwords.

In Figure 13, we present the percentage of men and women with respect to their number of
passwords. Almost the same percentage of men and women has one to four passwords. However,
26.53% of men and 41.18% of women have five to seven passwords. On the contrary, 26.53% of men
and 11.78% of women have eight or more passwords. So, we observe a significant difference
between men and women who use more than four passwords.

FIGURE 13: Number of passwords that men and women use
Taking into consideration the age of the users, we observe that the majority (76.19%) of people
between 46 and 60 years old use up to four passwords, while younger people tend to use many
passwords (Figure 14). Five to seven passwords are used by 35.59% of people between 18 and 30
years old, and by 40% of people between 31 and 45 years old. More than eight passwords are used
by 27.12% of people between 18 and 30 years old, and by 15% of people between 31 and 45 years
old.



FIGURE 14: Number of passwords that users have, according to their age.
It is also important to examine for what reasons they use those passwords, and how often they really
use them. Almost every user has a mobile phone (94%) and ATM cards (88%) with which he uses
passwords to verify his identity (Figure 15). Considering the increased use of Internet, we can
understand why so many people have passwords for their e-mail (64%), Internet connection (54%) or
to make a login to web pages (39%).

FIGURE 15: The applications where users use passwords.

Finally, we found that 69% of the participants use at least one password every day, and almost all of
them use a password once or twice per month. So, we can figure out how important is for everyone to
have a memorable and primarily safe password.

5.2 Analyzing text passwords
As we had already referred, text passwords nowadays are very unsafe and can cause many problems
to the people that use them. To find out if this situation is really true, we asked the participants what
kind of passwords they use for their applications. Their answers are presented in Figure 16.

FIGURE 16: Type of passwords that users choose.



The percentage of the users that choose passwords that are easy in memorization is 29%, while just
8% of them choose passwords that are difficult in memorization but also difficult to be cracked and as
a result very safe. The rest 63% of the participants chooses both easy and difficult passwords, based
on how important are the applications that are applied to.

These percentages are quite good at a first glance. But to make things more clear and to understand
if these results are really satisfactory, we have to find out more details about these passwords.

In Figure 17, we can see the number of characters that participants use in their passwords. 41% of
them have 4 characters in their passwords, 43% have 5 to 7 characters and 16% have more than 8
characters.

FIGURE 17: Number of characters in a text password.
Carefully looking at these results we can say that they are quite satisfactory as a password of more
than 8 characters is characterized very safe and really difficult to be cracked and one with 6 or 7
characters is characterized as a password of medium difficulty. But these results are incomplete,
because we have not examined yet what is the exact kind of the characters that these passwords are
consisted of. For this reason we have created Figure 18.

FIGURE 18: The exact kind of characters that the passwords are consisted of.

According to Figure 18, we can see that 49% of the participants use just numbers in their passwords
and 2% just letters. This situation is really problematic because most participants that use just
numbers or letters in their passwords create passwords with familiar dates or names that
consequently are very vulnerable to attacks. Only 15% of these passwords is consisted of numbers,
letters and symbols together and are really safe and reliable passwords.

5.3 New authentication techniques: Visual and Graphical passwords
Considering the previous results, we can be sure that most text passwords that were created by users
are predictable and really unsafe for them. So, it was inescapable that new authentication methods
were needed. These methods are visual and graphical passwords.



Next, after explaining to the users about visual and graphical passwords, we examine at what degree
they are positive in further learning and using them. Their answers are divided into five categories
(Figure 19), that are “by no means”, “a little”, “enough”, “much”, “very much”.

FIGURE 19: Users’ position in accepting or not of visual and graphical passwords

As we can observe in Figure 19, the users that are totally negative or a little interested in learning
more about the new authentication methods are very few (3% and 14% correspondingly). In
opposition to this, 36% of the users chose “enough” as their answer, 16% “much” and 31% “very
much”, a situation that reveals that most of the participants in our survey are positive in learning and
also using visual and graphical passwords.

5.4 Analyzing the three authentication methods
In this section we will analyze the two new authentication methods, visual and graphical passwords,
with respect to various parameters that will show us if these methods affected participants in a
positive way.

Firstly, we will examine the memorability of each method. As we can see in Figure 20, 57% of the
users believe that visual passwords are more memorable, while 43% of them prefer graphical
passwords.

FIGURE 20: Most memorable authentication method
Based on these results we found it very interesting to divide the participants into three categories,
describing their personality according to their personal opinion. To answer in this question the
participants had to select one out of the three choices that follow:
Visual, if the user remembers or learns something easier using images
Acoustic, if the user remembers or learns something easier by listening to it
Verbal, if the user remembers or learns something easier, when it is written down.
According to their answers, we conclude in Figure 21.



FIGURE 21: Most memorable authentication method based on the kind of person
that each user enlists himself.
Here we can see that more users that are visual types of persons prefer Visual Passwords (67.21%),
while users that are acoustic and verbal types of persons, prefer mostly Graphical Passwords
(56.25% and 60.87% correspondingly).

Next, we will examine and compare all three authentication methods (text, visual and graphical
passwords) regarding users’ opinion about how easy is it for them to learn how to use each method.
Users had to select one out of the five answers that were: “not at all”, “a little”, “enough”, “much” and
“very much”, in proportion to how difficult is the whole process for them. Collecting their answers, we
conclude in Figure 22.

FIGURE 22: How easy is for someone to learn how to use the three authentication
methods

According to Figure 22, we can observe that the users that found it very difficult or a little difficult to
learn how to use text and visual passwords are very few. Most of them found the two methods really
easy, for two different reasons:
Text passwords, because they are very familiar with them, as they use them almost every day
Visual passwords, because they are attractive and they have no special rules (the users just have to
choose a sequence of images as their password).

On the other side, only 2% of the users find it very easy to learn how graphical passwords are used
and 23% of them believe that the whole process is a little easy to be learnt. This percentage is not
really problematic, as this method is not so attractive at first side and has also many rules that a user
must remember, to create his password. What we have to mention here is that the rest of the
participants did not have difficulties in learning the use of graphical passwords and above all they
believe that with some practice everyone will be able to overcome any problem he may face.

The last parameter that we examined is how easy is for someone to use the three authentication
methods. Users here should again select one of the five given answers that are “not at all”, “a little”,



“enough”, “much” and “very much”, regarding how friendly they think that each method is for them.
The results are depicted in Figure 23.

FIGURE 23: User friendliness of the three methods

In Figure 23, we can see that text passwords are considered to be the friendliest method. For visual
and graphical passwords 5% and 2% correspondingly believe that they are not friendly at all to the
users, 20% and 27% that they are a little friendly, while the rest believe just the opposite. Moreover,
we have to say that 25% of the users chose “very much” as their answers for graphical passwords,
while the same percentage for visual is 10%. This situation proves us that even if visual passwords
impressed the users more at the beginning of the survey, the participants were able to overcome all
the difficulties and understand why graphical passwords is the best method out of the three.

5.5 Making more safe graphical passwords
After finishing processing the data, analyzing the results and taking into account other researchers’
results, we conclude that graphical passwords is the safest authentication method. Moreover with
some practice by the users, they can become very memorable and ease to be used by the users.

Considering all these, we suggest a set of tips and advices to users, in order to create very difficult
passwords and not vulnerable to attacks. These tips are the following:
do not make symmetrical shapes
do not make centered shapes
try to draw more than one lines
include in your drawing, as much pen up events as you can (the same drawing with more pen up
events is much safest)
avoid starting your drawing from the 4 corners (these cells are very vulnerable)
if your drawing is simple, try to make a second or even a third, similar to the initial one but covering
different cells
avoid drawing diagonal lines, and as a result lines near the intersection of the lines that create the
cells, because it is very easy to get confused and make shift errors.

6. CONCLUSIONS
We conducted our survey with the purpose to compare the proof by knowledge authentication
techniques which are text passwords, visual passwords and graphical passwords. A total of 100 users
participated in the survey. First, they were informed about the problems that text passwords have and
the advantages that visual and graphical passwords can offer to overcome these problems. Finally,
according to what they learnt and their personal opinion, they answered to our questionnaire.

Based on users’ answers we are able to confirm that almost every user uses passwords for different
applications such as mobile phone, ATM cards, credit cards, e-mails, internet connection etc. From
these users, almost half of men and women have 1 – 4 passwords, while about 40% of women have
5 – 7 and 26% of men 8 passwords or more. Indeed, as it was expected, users that have 1 – 4
passwords are the elder ones (46 – 60 years old) and those that have more passwords are the
younger.



Analyzing the text passwords that users claim that they create, we see that 30% of them create
passwords that are really memorable and as a result easy to be cracked. The encouraging thing is
that 63% of them create either easy or difficult passwords (with as many characters as they can), with
respect to the application that they are referred to. Unfortunately, this is not enough because many
users include only numbers in their passwords, or only letters, or numbers and letters both, creating
passwords of familiar dates and names that are vulnerable to dictionary attacks.

That was the main reason that a great percentage of users were very positive in knowing better the
new authentication methods. After learning more information about these methods, 57% of them
chose visual passwords as the most memorable method, while 43% preferred the graphical
passwords. Besides, as we divided the users in three categories, visual, acoustic and verbal, we must
say that visual users prefer mainly visual passwords while acoustic and verbal users prefer graphical
passwords.

Moreover, we must report that both new methods and especially visual passwords were characterized
very friendly by all users. At the same time graphical passwords were characterized a bit difficult until
they learn how exactly they are used.

Finally, keeping in mind that graphical passwords is the safest method, we made a small list with tips
that users must follow, in order to create really difficult graphical passwords and as a result very hard
to be cracked from anyone.

REFERENCES

[1] Bammigatti, P. H., and Rao, P. R. “Delegation in role based access control model for workflow
systems”. International Journal of Computer Science and Security, 2(2): 1-10, 2008.

[2] Chandrasekar, A., Rajasekar, V. R. and Vasudevan, V. “Improved authentication and key
agreement protocol using elliptic curve cryptography”. International Journal of Computer Science and
Security, 3(4): 325-333, 2009.

[3] Kar, J. and Banshidhar, M. “An efficient password security of multi-party key exchange protocol
based on ECDLP”. International Journal of Computer Science and Security, 3(5): 405-413, 2009.

[4] Tahir, M. N. “Hierarchies in contextual role-based access control model (C-RBAC)”. International
Journal of Computer Science and Security, 2(4): 28-42, 2008.

[5] Tahir, M. N. “Testing of contextual role-based access control model (C-RBAC)”. International
Journal of Computer Science and Security, 3(1): 62-75, 2009.

[6] W. Jansen. “Authenticating users on handheld devices”. In Proceedings of the Canadian
Information Technology Security Symposium, 2003.

[7] Bhagwat, R. and Kulkarni, A. (2010). “An overview of registration based and registration free
methods for cancelable fingerprint template”. International Journal of Computer Science and Security,
4(1): 23-30, 2010.

[8] H. Davies. “Physiognomic access control”. Information Security Monitor, 10(3): 5-8, 2005.

[9] K. Gilhooly. “Biometrics: Getting back to business”. Computerworld, May 2005.

[10] D. Klein. “Foiling the Cracker: a survey of, and improvements to, password security”. In
Proceedings of the 2n USENIX Security Workshop, pp. 5-14, 1990.



[11] de A. Angeli, L. Coventry, G. Johnson and K. Renaud. “Is a picture really worth a thousand
words? Exploring the feasibility of graphical authentication systems”. International Journal of Human-
Computer Studies, 63: 128-152, July 2005.

[12] H. Bolande. “Forget passwords, what about pictures?”. http://zdnet.com.com/2102-11-
525841.html

[13] X. Suo, Y. Zhu and S. G. Owen. “Graphical passwords: A survey”. In Proceedings of the Annual
Computer Security Applications Conference, Marriott University Park, Tucson, Arizona, 2005.

[14] K. Renaud and de A. Angeli. “My password is here! An investigation into visuo-spatial
authentication mechanisms”. Interacting with Computers, 16: 1017-1041, 2004.

[15] L. Sobrado and C. J. Birget. “Graphical passwords”. The Rutgers Scholar, 4, 2002.
http://RutgersScholar.rutgers.edu/volume04/contents.htm.

[16] F. Tari, A. A. Ozok and H. S. Holden “A comparison of perceived and real shoulder-surfing risks
between alphanumeric and graphical passwords”. In ACM International Conference Proceeding
Series, 149: 56-66, 2006.

[17] A. Perrig and D. Song. “Hash visualization: A new technique to improve real-world security”. In
Proceedings of the 1999 International Workshop on Cryptographic Techniques and E-Commerce
(CryTEC ’99).

[18] Real User Corporation. “About passfaces”, http://www.realuser.com/cgi-bin/ru.exe/_/homepages/
technology/passfaces.htm, accessed in November 2006.

[19] D. Davis, F. Monrose and M. Reiter. “On user choice in graphical password schemes”. In
Proceedings of the 13th USENIX Security Symposium, 2004.

[20] R. Dhamija and A. Perrig. “Déjà Vu: A user study using images for authentication”. In
Proceedings of the 9th USENIX Security Symposium, 2000.

[21] A. Bauer. “Gallery of random art”, 1998, http://andrej.com/art, accessed in December 2008.

[22] W. Jansen. “Authenticating mobile device users through image selection”. Data Security, May
2004.

[23] Passlogix. www.passlogix.com, accessed in November 2006.

[24] E. G. Blonder. “Graphical passwords”. Lucent Technologies, Inc., Murray Hill, NJ, U. S. Patent,
Ed. United States, 1996.

[25] S. Wiedenbeck, J. Waters, C. J. Birget, A. Brodskiy and N. Memon. “Authentication using
graphical passwords: Basic results”. In Human-Computer Interaction International (HCII 2005). Las
Vegas, NV, 2005.

[26] S. Wiedenbeck, J. Waters, C. J. Birget, A. Brodskiy and N. Memon. “Authentication using
graphical passwords: Effects of tolerance and image choice”. In Proceedings of the Symposium on
Usable Privacy and Security (SOUPS). Carnegie-Mellon University, Pittsburgh, 2005.



[27] S. Wiedenbeck, J. Waters, C.J. Birget, A. Brodskiy and N. Memon. “PassPoints: Design and
longitudinal evaluation of a graphical password system”. International Journal of Human Computer
Studies (Special Issue on HCI Research in Privacy and Security), 63: 102-127, 2005.

[28] I. Jermyn, A. Mayer, F. Monrose, K. M. Reiter and D. A. Rubin. “The design and analysis of
graphical passwords”. In Proceedings of the 8th USENIX Security Symposium. 1999.

[29] D. Nali and J. Thorpe. “Analysing user choice in graphical passwords”. Tech. Report TR-04-01,
School of Computer Science, Carleton University, Canada, 2004.

[30] C. P. van Oorschot and J. Thorpe. “On the security of graphical password schemes”. Technical
Report TR-05-11. Integration and extension of USENIX Security 2004 and ACSAC 2004 papers.

[31] J. Thorpe and P. Van Oorschot. “Graphical dictionaries and the memorable space of graphical
passwords”. In Proceedings of the 13th UNIX Security Symposium, August 2004.

[32] J. C. Birget, D. Hong and N. Memon. “Robust discretization with an application to graphical
passwords”. Cryptology ePrint Archive, Report 2003/168, http://eprint.iacr.org,

[33] K. Chalkias, A. Alexiadis and G. Stephanides. “A multi-grid graphical password scheme”. In
Proceedings of the 6th International Conference on Artificial Intelligence and Digital Communications,
Thessaloniki, Greece, 2006.

[34] A. Alexiadis, K. Chalkias and G. Stephanides. “Implementing a graphical password scheme that
uses nested grids”. In Proceedings of the International Conference for Internet Technology and
Secured Transactions (ICITST 2006), London, United Kingdom, 2006.

[35] I. Irakleous, M. S. Furnell, S. P. Dowland and M. Papadaki. “An experimental comparison of
secret-based user authentication technologies”. Information Management & Computer Security, 10:
100-108, 2002.

[36] B. Tribelhorn. “End user security”, 2002. http://www.cs.hmc.edu/~mike/public_html/courses/
security/s06/projects/index.html, accessed in November 2008.

[37] Y. Kim and T. Kwon. “An authentication scheme based upon face recognition for the mobile
environment”. In Proceedings of the International symposium on computational and information
science No1, Shanghai, China, 2004.

[38] J. Goldberg, J. Hagman and V. Sazawal. “Doodling our way to better authentication”. CHI '02
extended abstracts on Human Factors in Computer Systems, Minneapolis (ACM Press), 2002.

[39] R. Weiss and A. del Luca. “PassShapes: Utilizing stroke based authentication to increase
password memorability”. In Proceedings of the 5th Nordic conference on Human-computer
interaction: building bridges. Lund, Sweden, ACM pp. 383-392, 2008.

[40] A. M. Eljetlawi and N. Ithnin. “Graphical password: Prototype usability survey”. In Proceedings
IEEE International Conference on Advanced Computer Theory and Engineering, pp. 351-355, 2008.



[41] K. M. Everitt, T. Bragin, J. Fogarty and T. Kohno. “A comprehensive study of frequency,
interference, and training of multiple graphical passwords”. In Proceedings of the 27th international
conference on Human factors in computing systems. Boston, MA, USA. ACM, pp. 889-898, 2009.

[42] S. Chiasson, A. Forget, E. Stobert, P. C. van Oorschot and R. Biddle. “Multiple password
interference in text passwords and click-based graphical passwords”. ACM CCS’09, November 9–13,
2009, Chicago, Illinois, USA, 2009.

[43] A. A. Ozok and S. Holden “A strategy for increasing user acceptance of authentication systems:
Insights from an empirical study of user preferences and performance”. International Journal of
Business and Systems Research, 2(4): 343-364, 2008.

[44] K. Johnson and S. Werner. “Graphical user authentication: A comparative evaluation of
composite scene authentication vs. three competing graphical passcode systems”. Human Factors
and Ergonomics Society Annual Meeting Proceedings, 52: 542-546, 2008.

[45] M. D. Hafiz, A. H. Abdullah, N. Ithnin and H. K. Mammi. “Towards identifying usability and
security features of graphical password in knowledge based authentication technique”. In
Proceedings of the Second Asia International Conference on Modelling and Simulation, IEEE, pp.
396-403, 2008.

[46] L. Y. Por and X. T. Lin. “Multi-grid background Pass-Go”. WSEAS Transactions on Information
Science and Applications, 7(7): 1137-1148, 2008.


Aliaa A.Youssif, Ashraf A. Darwish & Ahmed Roshdy

Development of Information Agent Reranking By Using Weights
Measurement

Aliaa A.Youssif aliaay@helwan.edu.eg
Computer Science and Information Technology/
Computer Science/Helwan University
Helwan, Egypt

Ashraf A. Darwish amodarwish@yahoo.com
Computer Science/Mathematics/Helwan
University
Helwan, Egypt

Ahmed Roshdy ahmed.mind@hotmail.com
Computer Science/Mathematics/Helwan
University
Helwan, Egypt

Abstract

Web search is one of the most challenging problems of the Internet today,
seeking to provide users with search results most relevant to their information
needs. The new improvements of search engines technologies have made
available to the internet users an enormous amount of knowledge that can be
accessed in many different ways. However, there are some problems that face
the search engines. In this paper proposes an agent system which parses
information sources and extract weights that determine the powerful of relevant
information sources and prove that the word positions scores may affect on
reduction the relevance of those information sources. Moreover, it will show that
the user profile plays an important role in effectiveness of re-ranking and
updating ranking relevant web pages where agent learns user behavior by
observing user browsing for the interested result pages. In experimental work
section shows how weights algorithm gets more relevant web pages than other
algorithms that use word position which may reduce the value of relevance of
web pages.

Keywords: Information Agent, Knowledge-base, Weights, and Re-ranking.

1. INTRODUCTION
The vast amount information source in web today rendered the intelligent information agent
subtending to challenge re-ranking web pages on the fly. Given the constantly increasing
information overflow of the digital age, the importance of information retrieval has become critical.
Since the documents source of information retrieval have two types unstructured records and
semi-structured records depending on natural language text, and also kinds of data that can be
unstructured, e.g., photographic images, audio, video, etc [1,2].



The famous search engines on the marketing are now providing search facilities for databases
containing billions of web pages, where queries are executed instantly [3]. But there are some
problems that face search engines as following:

 The crawler-based search engines like Google can catalogues web pages and the
documents automatically where it crawls the web, then the users searches through what
they have found and sometimes they lead to poor queries and increase the gap between
information need and request [2].
 On the other side most of search engines beside Google crawls the web pages by
tracking hyperlinks to find authoritative pages using HITS algorithms that indeed get
satisfied relevant pages for users queries, but they suffer from another problem that
some authoritative pages have subjects and/or information sources and didn't satisfy the
relevance of the user request.

Though feedback is one approach that deduces information about the user and his/her search
context [4], beside that the techniques that depend on relevance feedback re-rank the search
results by re-calculating the relative importance of keywords in the query.

In this paper it concentrates on the processing of information sources and those contents to
extract words weights that are important to determine the content of the page subject more
powerful and can re-rank web pages more efficiently [4].

The paper is organized as follows: the next section is devoted to the related work that shows the
reasons of word weight importance more than its position that is used by the other agent's
techniques. Section 3 contains detailed description of the proposed information agent system,
showing its several components. The effectiveness of user information profile for learning agent
and effecting on re-ranking web pages in sections 4 and 5, are introduced respectively. The
report on an implementation and describe some experimental results in section 6 and conclusion
and future work have been presented in section 7.

2. RELATED WORK
The information agents developed and have addressed many tasks, such as assisting users in
searching and browsing the Web, finding, filtering and accessing large amounts of information on
behalf of users, and present a reduced and potentially relevant part of this information to them [5].
Information extraction from semi-structured documents has received more attention recently.
Agent in most cases is called PersonalSearcher [6] where it helps users to identify filtered
documents with high probability of being relevant to them by transforming the search process
according to users’ information preferences and learns about a user’s interests from the
observation of user browsing on the Web.

Some techniques focus on the structure of web page like CASA [1] and Textual Case-Base
Reasoning (TCBR) [7]. These algorithms depend on the position of the word either in text, line or
paragraph and give a degree for its position increasing from text till paragraph which in more
cases web pages puts words in text or low positions that refer to hyperlinks linked to addition
information that maybe more relevant to user request. Also the advertisements web pages
interest in the photos of goods and prices and little care about words positions.

Another technique in its algorithm calculate number of links that are linked to that page, and
checks position level of query keywords if it is either highest, middle, or lowest in web page as in
automated fuzzy agent [8]. Such this technique faces another problem that some pages have
titles and keywords of user query request and have high position but the information source is not
relevant for user query. Beside that some web pages have more relevant information sources and
titles either in medium or low position and add some commercial advertisements in upper layer. In
another hand the technique repeats itself for each page tracked by hyperlinks which make
overhead in processing.



The research focuses on extracting words weights to determine the rich information, density of
these words in the information source and how much they are more relevant to user query, and
they effect on re-ranking web pages neither depending on them position nor subject title position.

3. THE PROPOSED AGENT ARCHITECTURE
Since the agent is an integrated part of the end-program, there are two major paradigms for the
architecture. The first approach is automated treating with the results to get knowledge of certain
domain and learns from user or other agent for getting feedback about user interest. The second
approach is knowledge-based approach where the agent has extensive information of certain
domain and learns from user profile about his/her behavior and interest to expand the feedback of
the agent. The paper interests for the second approach and explain below of the architecture.

Figure 1 presents the high level view of the proposed architecture for information agent
depending on words weights and information profile. The knowledge-base system which is the
feedback of information agent has four major parts for processing documents as following:

(1) Entering information sources by downloading URLs and web pages that are retrieved from
the search engine or more search engines and indexing them for removing redundant URLs.
(2) Parsing web pages by parser for extracting semi-structured data, Meta tags and link tags
then determining words weights.
(3) Re-ranking web pages by calculating relevance equation ratio.
(4) Learning information agent from information profile about user browsing result web pages.

The user, by means of a graphical interface, submits a query to the parser by the keywords. The
parser extract query to its words and find word weights for each web page which is downloaded
in the knowledge-base. After that the knowledge-base calculate the relevance word weight ratio
to re-ranking web pages by descending and represent re-ranked result to user interface.

FIGURE 1: Show the high level of agent architecture

3.1. Knowledge-base
As it is explained in section 2 the importance of knowledge-base for information agent feedback
in certain domain. The architecture contains on the attributes of the domain, interactions, and
filtered results. The content of the base shown in Figure 2 is described as following:



DATA_EXTRACTION entity stores URL_ID (for not replication of URLs), URLs of result web
pages that are downloaded from the search engine, HTML source code, meta tags, hyperlink
tags, title and header tags, the number of queries that request a web page, and textfix that
contains paragraphs of data source and be processed from the parser.

LOCAL_WORDS entity stores the words of each page which is extracted by the parser and those
calculated weights. GENERAL_WORDS entity stores calculated word weight for all web pages
that are stored in the knowledge-base.

FIGURE 2: Show ER diagram of the knowledge-base

RERANKED_PAGE entity stores the calculated relevance word weight ratio for each page and
counting of user browsing for web pages result. REMOVED entity stores the characters, letters,
Preposition, etc that are used in paragraphs.

3.2. Parsing
The mission of the parser is extracting information sources of the web pages and determines the
kind of data source according to the requirements of search engine or information agent
technique [1, 9, 10, 11, 12]. The Paper focuses on the frequency of words in web pages to
determine those weights and find most relevant web pages depending on the concentration of
words query on those pages.

Therefore the more increasing of words query weights on a web page, the more determination of
relevant web page. Example for what to discuss about, when user submit specific query request,
the parser begins extract these words and compares them with those weights in knowledge-base
and then finds the highly weights that refer to highly relevant web pages after calculating
relevance word weight ratio.

4. USER PROFILE
Since user-profiling involves information about users of behavior-based information about them,
for example their actions on browsing some pages the which learns the information agent about
their interesting, besides increasing the information of the knowledge-base that are considered a
feedback of user interesting [13,14].

User behavior is the most deserving sources of information for getting profiles and, from their
successful interpretation; the information agent will be able to follow the user actions [15].
Therefore our agent monitors the user browsing for result pages and begins counting the user
behaviors by accessing web pages; the results will be automatically updated when unobserved



voting is highly score in some web pages than others. Implicit interest indicators like time
consumed in reading a Web page (considering its length), the amount of scrolling in a page, and
whether it was added to the list of bookmarks or not are considered a strong correlation with
explicit interest feedback, as opposed to the number of mouse clicks, which does not result in a
good indicator [16,17].

The dynamic part reflects the path of user behavior information in the way of user's search. The
static part lets user to show his/her desire when agent offers some options to the user. In many
cases user ignores static part [18].

5. RE-RANKING
Many search engines hide the mechanism used for the document ranking this is the reason of the
merging problem for the results becomes even more difficult. In addition, these kinds of
approaches suffer from ignoring or knowing nothing about the user conducting the search, nor the
context of the search [4].

Our evaluation for ranking of relative documents is depending on some weights using equation of
relevance which is discussed in [4]. For each document d in the response to query q, the
document rating is evaluated as follows (the adapted cosine function):

(1)

Where wd is the weight of word in the document d, wprof is the weight of chosen document d (i.e.
accessing web page from different users as mentioned in section 3), we supposed its score in the
first re-ranked iteration by 1, wq is the weight of query for chosen document from agent result, and
Wd is evaluated as following:

(2)

6. METHODOLOGY
The sequence of proposed algorithm can be obtained as following:
1) The user submits his/her search query.
2) The agent searches several other search engines with user’s query.
3) The agent downloads the resulted web pages
4) The agent extracts the information source from web pages using the Parser and saves
the information in the knowledge-base.
5) The agent calculates words weights of the query keywords.
6) The agent calculates relevance score for each web page and re-ranks them in a
descending order.
7) The agent keeps track of the re-ranked web pages. Once a user browses any of these
web pages, the wprof score of such page increases.
8) Steps 1-7 will be repeated for each search.
9) If the search hits any of the re-ranked web pages, it will increase its wq score.
10) The agent recalculates the relevance score and updates re-ranked web pages.

7. EXPERIMENTAL RESULTS
The experiment was constructed in three stages: training, retrieval and learning. For the former,
the knowledge-base was built using the method proposed in section 3.1 and determined the
domain to download web pages that was cars advertisements. Also parser was built to deal with



downloaded web pages for extracting information and re-ranking pages. In the retrieval stage,
knowledge-base was retrieved with each first few web pages were downloaded from five search
engines (Google, AltaVista, Yahoo, MSN, and Excite) and indexed these pages to prevent
redundant pages. In the program we generated 20 different queries and calculated summation
average of number of relevant pages and irrelevant pages. In learning stage we browsed in re-
ranked pages to learn information agent of interested pages by accessing them and calculating
period time of browsing for each page; the longest time taken page get score in wprof and be
affected for re-ranking again. The stages are described briefly as following:

7.1. Training Stage:
(1) Build a knowledge-base and determine the domain for downloading pages.
(2) Build parser.

7.2. Retrieval Stage:
(1) Download first 200 web pages from each search engine.
(2) Index web pages to prevent redundant web pages.
(3) Generate different queries and calculated summation average.

7.3. Learning Stage:
For each term, Browsed in re-ranked pages to give interested pages scores in wprof for next re-
ranking.

The Figure 4 represents the most common words of user query that insure the relevance of web
page.

FIGURE 3: Show Agent Search Result

7.4. Comparative Analysis
In order to evaluate our proposed algorithm, we compared our findings with the score resulted
from TCBR [7], CASA [1] Automated fuzzy algorithm [8]. Table 1 shows the comparative
evaluation of common queries that have effect of word position in average of first 20 relevant web
pages.



Automated CASA Proposed
fuzzy score
TCBR score score

MAZDA ph 3.016 2.255 0.918 4.565

FIAT Audio 3.927 3.573 1.641 5.126

BMW Engine 4.537 3.847 2.177 6.337

HONDA in
Boston
2.472 1.501 0.861 3.691

Toyota
4.582 3.674 2.255 5.969
Accessories
PROTON-GN2
Weight
3.415 2.293 1.106 4.846

PEUGEOT
Spares
4.357 3.063 2.144 5.138

JEEP
2.732 1.587 0.161 3.813
Cylinder

CHERRY 3.157 2.059 0.345 4.152

NISSAN
Carpet
3.437 2.462 1.147 4.598

Used
3.843 2.826 1.164 5.326
LANOS
Mercedes
Brake
2.952 2.147 1.030 4.293

SEAT Filter 2.638 1.476 0.185 3.232

Hyundai
2.431 1.630 0.289 3.813
Lights
SKODA
Hatchback
4.362 3.326 2.577 6.429

TABLE 1: The difference relevant scores between algorithms

From Table 1, we noticed that our proposed algorithm gives higher word score than CASA, TCBR
and fuzzy algorithm. In computing the relevance of a web page, our equation does not take into
account the word location.

However, in some queries like “Toyota Accessories”, “Proton-GN2 Weight”, “Honda in Boston”
the factor of word location in CASA algorithm and ∆ij in TCBR equation for word weight
weight(wj,di) = tfij + ∆ij decrease (where tfij is term frequency that determines the frequency of word
appearance in document) because these words occurred more frequently in lines than in
paragraphs and titles.

Moreover, in the automated fuzzy algorithm we noticed that the Position_Score of some words
decreases to some extent. This is due to the fact that these words occurred in the middle of the
web pages, which affected their score. The relevant web page score is determined by the
equation:
Score = (2*Frequency_Score) + Position_Score + Links_Score

Summary:
Figure 3 shows the comparison results between 3 algorithms: CASA algorithm [1], TCBR
algorithm [7], automated fuzzy algorithm [8] and weights measures algorithm of the first 200 re-
ranked web pages from information agent and that describes the performance of agent can be
more efficient depending on weights measurement.



200
180
Relevant web pages
160
No. of retrieved 140
Irrelevant web pages
web pages 120
100
80
60
40
20
0
proposed CASA automated TCBR
algorithm algorithm fuzzy algorithm algorithm

FIGURE 4: Comparison between four algorithms

The performance of information agent can be increased by improving the mechanism of
downloading URLs and web pages using semantic web techniques as in [11,12,19] that
determine word vectors of context meaning , and reaching to suitable solution to the retrieve
relevant documents dynamically.

8. CONCLUSION AND FUTURE WORK
In this paper we scoped on filtering powerful information source by parsing the information to
extract weights for finding more relevant pages and re-ranking web pages with learning
information agent of user profile about interesting web pages that can be more efficient in re-
ranking. The result of the experiment for proposed algorithm is 0.82% of relevant web pages
compared to highly percentage of other algorithms about 0.71%. We used Microsoft Access to
build knowledge-base and Visual Basic to build the other components of information agent.

9. REFERENCES
[1] X. Gao, "A methodology for building information agents", in Yang, Y., Li, M., Ellis, A. (Eds),
Web Technologies and Applications, International Academic Publishers,
<http://www.cs.mu.z.au/~xga/apweb/index.htm >, pp.43-52. (1998)
[2] Ed Greengrass "Information Retrieval: A Survey", Available at: http://www. csee.umbc.edu.
(2000)
[3] M. Caramia, G. Felici, and A. Pezzoli "Improving search results with data mining in a
thematic search engine", Computers and Operations Research Volume 31, Pages: 2387 –
2404, ISSN: 0305-0548. (2004)
[4] B. Chidlovskii, N.S. Glance and M.A. Grasso, "Collaborative Re-Ranking of Search Results",
Available at: http://citeseer.ist.psu.edu. (2000)
[5] Daniela Godoy, Silvia Schiaffino, Analia Amandi "Interface agents personalizing Web-based
tasks", Cognitive Systems Research Volume 5, Issue 3, Pages 207-222. (2004)
[6] Godoy, D., & Amandi, A. "Personalsearcher: An intelligent agent for searching web pages" In
International joint conference IBERAMIASBIA’2000 (pp. 43–52). Atibaia, Sao Paulo, Brazil:
Springer. (2000)
[7] D. Lis Godoy, "Generating User Profiles for Information Agents", Available at:
http://citeseer.ist.psu.edu. (2003)
[8] M. Mohammadian in the book "Intelligent Agents for Data Mining and Information Retrieval",
Idea Group Publishing (an imprint of Idea Group Inc.), ISBN: 1-59140-277-8. (2004)
[9] L. Shen and A.K. Joshi, "An SVM Based Voting Algorithm with Application to Parse
Reranking", the seventh conference on Natural language learning at HLT-NAACL 2003 -
Volume 4, Pages: 9 - 16. (2003)
[10] A. Penev and R. Wong, "Shallow NLP techniques for Internet Search", the 29th Australasian
Computer Science Conference - Volume 48, Pages: 167 – 176, ISBN ~ ISSN: 1445-1336, 1-
920682-30-9. (2006)



[11] C. Cesarano, A. d’Acierno and A. Picariello, "An Intelligent Search Agent System for
Semantic Information Retrieval on the Internet", the 5th ACM international workshop on Web
information and data management, Pages: 111 – 117, ISBN:1-58113-725-7. (2003)
[12] G. Wisniewski & P. Gallinari "From Layout to Semantic: A Reranking Model for Mapping Web
Documents to Mediated XML Representations" proceeding of RIAOCID (2007) Conference,
URL: http://dblp.uni-trier.de/db/conf/riao2007.html#WisniewskiG07. (2007)
[13] Bas van Gils and Eric D., "User-profiles for Information Retrieval", Methodologies for
Intelligent Systems, volume 542 pages 102-111 ISBN: 978-3-540-54563-7. (2006)
[14] S.E. Middleton, "Capturing knowledge of user preferences with recommender systems",
Available at: http://citeseer.ist.psu.edu. (2003)
[15] D. Godoy and A. Amandi "User profiling in personal information agents: a survey", The
Knowledge Engineering Review, Volume 20, Pages: 329 - 361 ISSN: 0269-8889. (2005)
[16] Claypool, M., Le, P., Wased, M., & Brown, D. "Implicit interest indicators. Intelligent User
Interfaces", Proceedings of the 6th international conference on Intelligent user interfaces
Santa Fe, New Mexico, United States Pages: 33 – 40 ISBN:1-58113-325-1 (2001)
[17] Kelly, D., & Belkin, N. J. "Reading time, scrolling and interaction: Exploring implicit sources of
user preferences for relevant feedback", proceedings of the 24th annual international ACM
SIGIR conference on research and development in information retrieval (pp. 408–409). New
Orleans, LA, USA: ACM Press. (2001)
[18] C. Dharap "Context-based and user-profile driven information retrieval", Fremont CA (US),
Philips Corporation New York (2001).
[19] G. Cheng, W. Ge and H. Wu, "Searching Semantic Web Objects Based on Class
Hierarchies", proceeding of LDOW Conference, available at:
http://iws.seu.edu.cn/publications/cgwq08 . (2008)


International Journal of Computer Science and Security Volume (4) Issue (2)

More Related Content

What's hot (18)

Viewers also liked (10)

Similar to International Journal of Computer Science and Security Volume (4) Issue (2) (20)

International Journal of Computer Science and Security Volume (4) Issue (2)