Using Learning Vector Quantization in IDS Alert Management System

Amir Azimi Alasti Ahrabi, Kaveh Feyzi, Zahra Atashbar Orang, Hadi Bahrbegi & Elnaz Safarzadeh
International Journal of Computer Science and Security, (IJCSS), Volume (6) : Issue (2) : 2012 128
Using Learning Vector Quantization in Alert Management of
Intrusion Detection System
Amir Azimi Alasti Ahrabi amir.azimi.alasti@gmail.com
Department of Computer
Islamic Azad University, Shabestar Branch
Shabestar, East Azerbaijan, Iran
Kaveh Feyzi kavehfeizi@gmail.com
Ataturk University
Erzurum, Turkey
Zahra Atashbar Orang atashbarorang_z@yahoo.com
Islamic Azad University, Tabriz Branch
Tabriz, East Azerbaijan, Iran
Hadi Bahrbegi hadi.bahrbegi@gmail.com
Elnaz Safarzadeh elnaz_safarzadeh@yahoo.com
Abstract
Intrusion detection system (IDS) is used to produce security alerts to discover attacks against
protected network and/or computer systems. IDSs generate high amount of security alerts and
analyzing these alert by a security expert are time consuming and error pron. IDS alert
management system are used to manage generated alerts and classify true positive and false
positives alert. This paper represents an IDS alert management system that uses learning vector
quantization technique to classify generated alerts. Because of low classification time per each
alert, the system also could be used in active alert management systems.
Keywords: IDS, Alert Management, Learning Vector Quantization, Alert Classification, True Positive and
False Positive Classification.
1. INTRODUCTION
An intrusion detection system (IDS) inspects all inbound and outbound network activity or
computer system events and identifies suspicious patterns that may indicate a network or system
attack from someone attempting to break into or compromise a system. [1]. IDSs are producing
many alerts each day that many of them are false positive alerts. Big amount of the false positive
alerts crowd and cover true positive alerts from security experts. Also identifying true positive
from false positives are time consuming and error prone therefore IDS alert management system
are introduced to manage generated IDS alerts. IDSs can be used as active or passive. In
passive usage of IDS, it analyzes traffics or events in offline mode but active IDSs work in online
mode. To manage alerts concurrently with alerts generation, active alert management systems
are used. Active alert management systems same work in online mode as active IDSs. These

types of alert management systems should have little amount of alert analyze time to be used in
online mode. Some of problems of IDS are: huge amount of generated alerts and high rate of
false positive alert among generated alerts. Also most alert management system has low speed.
In this paper authors change their previous work and proposed a new alert management system
by using Learning Vector Quantization (LVQ) [2]. It classifies the generated alerts based on
attack type of alerts, detects false positive alerts, high speed classification to use with alert
generation in IDSs. The proposed system uses some techniques of previous work techniques [3]
such as alert filtering, alert preprocessing, and alert filtering to improve accuracy of the results.
In Section 1 the alert management system is introduced. Section 2 reviews related works, section
3 explains the suggested alert management system and describes all component of the proposed
system, the experimental results are shown in section 4 and finally section 5 is a conclusion and
future works.
2. RELATED WORKS
Alert management systems use various method and techniques. Clustering and classification of
alerts is one of these techniques. A method of clustering based on root causes is proposed by K.
Julisch [4] which clusters IDS alerts by discovering main cause of their occurrences. He proves
that a small number of root causes imply 90% of alerts. By removing alerts related with these root
causes total number of alerts come down to 82%. The system uses information about underlying
network so it is not portable and this problem is a disadvantage of the algorithm.
Three artificial intelligence techniques with some dimension reduction techniques are used to
cluster generated IDS alerts from DARPA 2000 dataset in [5] then produced results are
compared. The problems of that system are: row alert without preprocessing are entered to the
algorithms and system is not tuned. Cuppens proposed another method that uses expert system
to make decision [8, 16]. In [6, 7] two genetic clustering algorithm based, named Genetic
Algorithm (GA) and Immune based Genetic Algorithm (IGA) used to manage IDS alerts. Their
proposed methods depend on underlying network information same as method proposed by
Julisch.
Wespi et al. [17] design a system that aggregates alerts together by placing them in situations.
Situations are set of special alerts. To construct a situation, source, destination and attack class
attributes of alert are used.
Authors of this paper propose a system that manages alert generated from DARPA 98 dataset
[3]. Some algorithms such as alert filtering, alert preprocessing and cluster merging are used in
the system. The main unit of the system is cluster/classify unit that uses Self-Organizing Maps
(SOM) [2] to cluster and classify IDS alerts. Results of [3] show that SOM was able to cluster and
classify true positive and false positive alerts more accurate than other techniques.
In another work, authors have developed an alert management system [9] similar to [3]. In that
work usage of seven genetic clustering algorithms named Genetic Algorithm (GA) [18], Genetic
K-means Algorithm (GKA) [19], Improved Genetic Algorithm (IGA) [20], Fast Genetic K-means
Algorithm (FGKA) [21], Genetic Fuzzy C-means Algorithm (GFCMA) [22], Genetic Possibilistic C-
Means Algorithm (GPCMA) [9] and Genetic Fuzzy Possibilistic C-Means Algorithm (GFPCMA) [9]
to cluster and classify true positive and false positive alerts, are explained. The system after
clustering alerts then prioritized produced clusters with Fuzzy Inference System [9].
In this paper an alert management system based on system proposed by authors in [3] is
proposed that uses LVQ as a tool to classify input alert vectors. Propose of this paper evaluating
another type of Kohonen networks named LVQ [2] in alert management system field. The system
will be able to improve accuracy of results and also to reduce the number of false positive alerts.

3. USING LVQ IN ALERT MANAGEMENT SYSTEM
The proposed system is shown in Figure 1. In this paper we use binary traffics files of a network
named DARPA 98 dataset [10] instead of real network traffics. Snort tool [11] is used to produce
alerts of DARPA 98 dataset network traffics. Snort is an open source signature based IDS which
gets DARPA 98 online traffic and then generates alert log files [3]. After generating alert log files
with Snort tool, these files are entered to the proposed system as its input.
FIGURE 1: Proposed alert management system.
3.1 Labeling Unit
Labeling unit gets generated alert from Snort and tcpdump.list files of DARPA 98 dataset and
then generate labeled alert which each alert has own attack type. tcpdump.list files contain
information about all packets in DARPA 98 dataset. These labels are used to train LVQ and
evaluate results of LVQ [3, 9].
3.2 Normalization and Filtering Unit
In this phase accepted attack types are entered to the unit and only alerts that are in class of
these attack types are selected [3, 9,12]. This unit uses eight attributes of alert to filter alert, this
attributes are: Signature ID, Signature Rev, Source IP, Destination IP, Source Port, Destination
Port, Datagram length and Protocol [13]
3.3 Preprocessing Unit
Preprocessing unit converts string values of attributes of alert to numerical data. It also reduces
the range of attribute values and converts alerts to data vectors (1), (2) and (3).
4321
4321
255)255))255(((_
,...
XXXXVALIP
XXXXIP
+×+×+×=
=
(1)







=
=
=
=
=
UDPprotocol
TCPprotocol
ICMPprotocol
Noneprotocol
valprotocol
,17
,10
,4
,0
_ (2)

1.08.0
minmax
min
+
−
−
×=
xx
xx
IUR (3)
3.4 LVQ Training and Classification Unit
In this unit we use LVQ as a classifier. LVQ should be trained with train dataset and then gets test
dataset to classify them.
• Learning Vector Quantization
LVQ is a special artificial neural network; it applies a winner-take-all Hebbian learning-based
approach. LVQ was invented by Teuvo Kohonen. It is a forerunner to SOM and related to Neural
gas, and to the k-Nearest Neighbor algorithm (k-NN) [2].
An LVQ system is represented by prototypes W=(w(i),...,w(n)) which are defined in the feature
space of data vectors. In winner-take-all training algorithms, the prototype which is closest to the
input vector according to a given distance measure for each vectors of input data are determined.
The position of this so-called winner prototype is then adapted, i.e. the winner is moved closer if it
correctly classifies the data point or moved away if it classifies the data point incorrectly.
4. EXPERIMENTAL RESULTS
To simulate the proposed system C#.net programming language, MATLAB software and SOM
toolbox is used [14, 15]. The parameters of simulation are shown below.
Suggested LVQ has 80 neurons in hidden layers. The LVQ gets a data vector of train data that
each data vector consists of 8 attributes as input to the system. Training phase consists of 50
epochs. Learning function is learnlv2. Because of Input data vectors consist of 9 alert attack
types, each attack type have typical class percentage 0.1 except false positive. False positive
typical class percentage is 0.2. The attack types used in this simulation are: Back, Pod, Nmap,
Imanp, Dict, Rootkit, Land and Phf. Train data contains 70% of total filtered alert data vectors or
10166 data vectors. The false positive count in the training dataset is 4113. Test dataset includes
30% of the data vectors of labeled alerts; it means 2591 data vectors of true positive, and 1764
data vectors of false positive alerts.
Figure 2 shows Mean Square Error (MSE) for each epoch. As you can see in this figure the error
value is reduced when we moved forward on epoch axis; and minimum value of the error
achieved in last step.
To evaluate the performance of algorithms four measurements are introduced, they are:
1- Classfication Error (ClaE),
2- Classfication Accuracy percent (ClaAR),
3- Average Alert Classification Time (AACT),
4- False Positive Reduction Rate (FPRR).
In table 1 value of these metrics are shown. The values of ClaE and ClaAR are 490 and 88.75%
respectively (Table 1). The value of AACT measurement is 0.000018 that shows the proposed
system can be used in active IDS alert management systems that evaluate alerts while IDS
produces them simultaneously. False positive alert type identification known FPRR is an
important point of extracted values. Because of production of false positive alerts beside true
positive ones then this metric value is very important in modern IDS alert management systems.
The value of this metric is 88.27% percent.

FIGURE 2: Errors of ANN output values per targets.
AACTFPRRClaARClaE
0.00001888.2788.75490
TABLE 1: Extracted performance metric values from simulation.
In [9] GA based algorithms are used to cluster and classify alerts. These results are shown in
table 2. For ClaE, ClaAR and FPRR metrics the proposed system has high value in contrast of
GA and GKA. But other methods such as IGA, FGKA, GFCMA, GPCMA and GFPCMA have
better performance in contrast proposed system. In AACT performance metric, LVQ based alert
management system has better result than all of GA based techniques. It means that LVQ could
be used in active alert management system.
Algorithm ClaE ClaAR FPRR AACT
GA 1218 72.03 52.15 Offline
GKA 1011 75.2 62.11 Offline
IGA 306 92.97 95.24 Offline
FGKA 314 92.79 97.51 Offline
GFCMA 148 96.60 97.51 Offline
GPCMA 91 97.91 96.03 Offline
GFPCMA 148 96.60 97.51 Offline
TABLE 2: Results of performance metrics for GA based algorithms.
5. CONCLUSION AND FUTURE WORKS
In this paper a LVQ based system is presented that is able to classify IDS alerts. The system
solved some problems of IDSs such as generating high amount of alerts and false positive alert.
The system could classify true positive alert and could identify false positive ones. The system
identifies and drastically reduces the number of false positive alerts. The results of the proposed
system are compared to GA based techniques. The comparison shows that in contrast of GA
based systems LVQ algorithm can be used in active alert management systems.
It seems to be useful using LVQ to correlate alerts to discover attack sequences so this idea is
another future work of this paper.

6. REFERENCES
[1] H. Debar, M. Dacier, and A. Wespi. "Towards a taxonomy of intrusion-detection systems”,
COMPUT. NETWORKS, Vol. 31, Issue: 8, pp.: 805-822, 1999.
[2] Kohonen, T, "Self-Organized Maps", Springer series in information. Science Berlin
Heidelberg, 1997.
[3] Amir Azimi Alasti Ahrabi, Ahmad Habibizad Navin, Hadi Bahrbegi, Mir Kamal Mirnia, Mehdi
Bahrbegi, Elnaz Safarzadeh, Ali Ebrahimi, "A New System for Clustering and Classification of
Intrusion Detection System Alerts Using Self-Organizing Maps", International Journal of
Computer Science and Security (IJCSS), Vol. 4, Issue 6, pp. 589 – 597, 2010.
[4] K. Julisch, "Clustering intrusion detection alarms to support root cause analysis", ACM Trans.
on Information and System Security, Vol. 6, Issue 4, pp. 443 – 471, 2003.
[5] Maheyzah, M. S., Mohd Aizaini, M., and Siti Zaiton, M. H. (2009), "Intelligent Alert Clustering
Model for Network Intrusion Analysis", Int. Jurnal in Advances Soft Computing and Its
Applications (IJASCA), Vol. 1, Issue 1, pp. 33 – 48, 2009.
[6] Wang, J., Wang, H., Zhao, G., "A GA-based Solution to an NP-hard Problem of Clustering
Security Events", IEEE, pp. 2093- 2097, 2006.
[7] Wang J., Baojiang Cui, "Clustering IDS Alarms with an IGA-based Approach", ICCCAS, pp.
586-591, 2009.
[8] Cuppens F., “Managing alerts in a multi-intrusion detection environment”, Proceedings of the
17th Annual Computer Security Applications Conference on, pp. 22-31, 2001.
[9] Bahrbegi H., Navin A.H., Ahrabi A.A.A., Mirnia M. K., Mollanejad A., "A new system to
evaluate GA-based clustering algorithms in Intrusion Detection alert management system",
Nature and Biologically Inspired Computing (NaBIC), Second World Congress on, pp. 115 –
120, 2010.
[10]MIT Lincoln Lab., DARPA 1998 Intrusion Detection Evaluation Datasets. Available:
http://www.ll.mit.edu/mission/communications/ist/corpora/ideval/data/1998data.html, 1998.
[11]Snort: The open source network intrusion detection system. Available: http://www.snort.org/.
[12]Brugger S. T., J. Chow, "An Assessment of the DARPA IDS Evaluation Dataset Using Snort",
UC Davis Technical Report CSE-2007-1, Davis, CA, 2007.
[13]Snort Manual, www.snort.org/assets/82/snort_manual.pdf.
[14]Neural Network Toolbox, "ANN Toolbox for MATLAB", www.mathworks.com/products/neural-
network, 2011.
[15]Matlab Software, http://www.mathworks.com.
[16]E. MIRADOR, "Mirador: a cooperative approach of IDS", European Symposium on Research
in Computer Security (ESORICS). Toulouse, France, 2000.
[17]Debar H., Wespi A., "Aggregation and Correlation of Intrusion-Detection Alerts", Proceeding
RAID '00 Proceedings of the 4th International Symposium on Recent Advances in Intrusion
Detection, pp.:87-105, 2001.

[18]Krovi R., "Genetic Algorithm for Clustering: A preliminary investigation", Proceeding on 25th
Hawaii International Conference on Systems Sciences (HICSS), pp. 540–544, 1992.
[19]Krishna K., Murty M., "Genetic K-means algorithm", IEEE Transactions on Systems, Man and
Cybernetics - Part B: Cybernetics, pp. 433-439, 1999.
[20]Fuyan L., Chouyong C., Shaoyi L., "An Improved Genetic Approach", International
Conference on Neural Networks and Brain, pp. 641-644, 2005.
[21]Lu Y., Lu S., Fotouhi F., Deng Y., Brown J. S., "FGKA: a Fast Genetic K-means Clustering
Algorithm", Proceeding of the ACM Symposium on Applied computing (SAC), Nicosia,
Cyprus, pp. 622-623, 2004.
[22]Nuovo A. D. G., Catania V., Palesi M., "The Hybrid Genetic Fuzzy C-means: a Reasoned
Implementation", Proceedings of the 7th WSEAS International Conference on Fuzzy
Systems, ACM, pp. 33-38, 2006.

Using Learning Vector Quantization in IDS Alert Management System

More Related Content

What's hot (17)

Similar to Using Learning Vector Quantization in IDS Alert Management System (20)

Recently uploaded (20)

Using Learning Vector Quantization in IDS Alert Management System