NETWORK INTRUSION DATASETS USED IN NETWORK SECURITY EDUCATION

International Journal on Integrating Technology in Education (IJITE) Vol.7, No.3, September 2018
DOI :10.5121/ijite.2018.7304 43
NETWORK INTRUSION DATASETS USED IN
NETWORK SECURITY EDUCATION
Alia Yahia1
and Eric Atwell2
1
Sudan University of Science and Technology, Khartoum, Sudan
2
University of Leeds, Leeds, UK
ABSTRACT
There is a gap between the network security graduate and the professional life. In this paper we discussed
the different types of network intrusion dataset and then we highlighted the fact that any student can easily
create a network intrusion dataset that is representative of the network they are in. Intrusions can be in
form of anomaly or network signature; the students cannot grasp all types but they have to have the ability
to detect malicious packets within his network.
1. INTRODUCTION
The need to gap between theoretical and practical network security is a must as black hats are
getting more aggressive by the day and the white hats must keep up with new attacks. Student
learning network intrusion should be able to see attack signatures and learn the different
techniques to detect them. The data used in generating these datasets; should it be live or
simulated, should be labeled putting into account signature base or anomaly base detection to
detect new attacks. The two mostly wide datasets in network intrusion research are the KDD
CUP99 and DARPA 1998-1999 are still in use although they face strong criticism. Laskov
highlighted the reasons why these datasets are used in spitecriticism generating new datasets with
labeled packets, which is extremely time consuming and sometimes impossible.Live data does
not cover all types of attacks available which hinders the training process.
2. DATASETS
2.1 DARPA
DARPA network dataset intrusion has been created for the purpose of evaluating the different
Algorithms used to detect intrusions (Ciza et al., 2008). It has been criticized for its lack of ability
to detect zero base attacks and the absence of false positives (Zuechet al., 2015). The data
generated is made up of backward data that is intended to be completely free of attacks and attack
data that is intended to consist entirely of attack scenarios. It is argued by McHugh (2000)that one
can really simulate data that mimics the live data since it isunpredictable and does not follow a
certain norm or pattern. The 300 attacks where synthesis in the data, but the order of the attacks
was consistence through the 10 weeks of capture which can be argued as unrealistic. The purpose
of the educational dataset is to help students understand intrusions and their detection and
DARPA lacks this support due to the attack taxonomy used. (McHugh, 2000)
2.2 KDD
KDD99 dataset is the most widely used dataset when combining the domains of network
intrusions with machine learning (Ozg ̈ur& Erdem, 2016);despite the fact that there are redundant

44
records in the dataset which makes the detection process skewed and unreliable. (Zuech et al.,
2015).“Dataset contains 24 attack types in training and 14 more attack types in testing for total of
38 attacks. These 14 new attacks theoretically test IDS capability to generalize to unknown
attacks. At the same time, it is hard for machine learning based IDS to detect these 14 new
attacks” (Ozg ̈ur& Erdem, 2016). Thus the main objective of the KDD is not achieved; which is
to help in the evaluation of machine learning algorithms.
2.3 INDIAN RIVER STATE COLLEGE DATASET
The dataset generated is a hybrid between IRSC dataset representing real data and simulated
network attacks generated by the network team. (Zuech et al., 2015) Full packet capture and
Network Flow data capture are used. Full packet capture are achieved by capturing the packets
using Snort. Data cleaning is performed by capturing the packets using both Snort and Wireshark,
captured are compared and any missing packets merged. Network Flow data capture provide a
higher level of abstraction since the volume of data it collected is less and it is summarized.
(Zuech et al., 2015) Labeling simulated attack is straightforward because the network teach are
the ones who generated the attacks. Labeling real attacks is done in a multistage process, stage 1
use Snort rules to detect intrusions and then pass them to the network flow component. Manual
inspection is then performed. The advantage of this dataset it that it captures packets from a live
network which reflects real traffic scenarios which is one of the problems of the DARPRA
dataset.
2.4 KYOTO 2006+
Kyoto 2006+ is an evaluation dataset where the data is obtained from diverse honeypots from
November 2006 to August 2006. One drawbacks is that the data is mostly intrusive which make
the training method skewed. The dataset is labeled where 1(normal session) , -1 (known attacks)
and -2( unknown attacks) . The dataset does not contain any attacks generated from or targeting
windows machines which is a problem since the number 1 desktop operating system. The number
of protocol types is limited to TCP , UDP and ICMP which does not reflect the different types of
attacks found.
3. PROBLEM DEFINITION
Creating a network intrusion detection educational corpus using corpus linguistics methodology
with a limited number of features and from real production network will led to a balance corpus
that will contain different types of intrusions that can form the basis in which students can learn
cyber security.
4. METHODOLOGY
Merging the corpus linguistics methodology with network security model will lead to a network
dataset that can be used as a tool in teaching the different types of network attacks found in real
network datasets. It will contain packets that are representative of real life scenarios.
4.1 CORPUS LINGUISTICS METHODOLOGY AND NETWORK SECURITY METHODOLOGY
TCP/IP is the language of the Internet. Its main objective was to create a friendly communication
between two heterogeneous entities. Over time it has been abused to allow intrusive

45
communication. This has been possible by the vulnerabilities found in all aspects of the
communication process from protocols,hardware, system software and application software.
“A corpus is a collection of machine-readable texts that have been produced in a natural
communicative setting. They have been sampled to be representative and balanced with respect to
particular factor.” (Evans, 2009). The network security model cycle is applied to build a corpus of
intrusions .The cycle consist of collection,detection and then analysis.
Collection can be full content data, session data, statistical data , packet string data and alert data .
Detection according to intrusion detection system can be classified into vulnerability centric and
threat centric (Ciza et al., 2008) Threat centric was adopted because it is base on collection and
relies on utilizing all sources of attack .
Analysis can be packet analysis , network forensics , host forensics and malware analysis.
4.2 COLLECTION
Before collecting the data we designed the corpus.There were a number of parameters that had to
be addressed before creating the corpus (Evans, 2009). The size of the corpus(dataset) should
represent the different type of intrusions so the students can learn from it. This will not be visible
but at least there should be an example of a network attack for each of the protocol dialects
represented in this paper by the major protocols found in each of the TCP/IP protocol stack.
4.3 BALANCE
We have collected the dataset from different networks at different time and days. Hopefully this
will bring balance. The people (the computers. mobiles and processes) communicating using
TCP/IP where diverse using different operating system and hardware architectures.
4.4 REPRESENTATIVENESS
The packet capture were captured from inside the Sudan network , and our domain is to teach
students inside Sudanese university the different type of network intrusions found inside
Sudanese networks.
The capture was done using Wireshark ; which is an opens source packet sniffer.
4.5 ANALYSIS AND DETECTION
The second research question was can we label the packets as intrusive or not by using Wireshark
as network intrusion dataset. We have adopted network base,signature base detection so the
percentage of false negative becomes zero. This was not possible since no baseline of normal
activities was created and network usage. So despite the fact that a signature can appear as an
intrusion e.g the data value in the protocol hierarchy statistics; which means that Wireshark has
located a dialect (protocol) it cannot detect , this can be an intrusion or a new application in the
market and Wireshark has not built a protocol dissector for it.
4.6 EXPERIMENT
Intrusion detection starts with where to place the sensor that will capture the packets, in this
situation, where to place our sensors was overlooked as it was one of the research question is it
really of importance. We captured 11 datasets from different sites on different days, our first
objective was to merge them but this was not possible because they used the same private address

46
space which would haveled to duplicate and usually the attack is made of a stream of packets, if
merge the stream will be invalid.
We adopted the five classes of attacks found DARPRA to see if they are present or not. One of
the limitation DARPRA was that it did not represent real present attacks so if our datasets did not
have at least one type of the major type of attack then it falls short of being fit as a network
intrusion educational dataset.
Classes of attacks are
• Denial of service: looked for the SYN flood , ICMP flood , UDP flood (Degadzor et al.,
2017) as a method of detection also Time To Live expiry attack. Found this attack in the
10 of 11 datasets.
• Remote to user and user to root where not present, here we searched for brute force
attacks
• Surveillance attacks : found in all the sweep attacks of different types.
Other attacks
• Arp poisoning method of detection was the expert info tab where the warning duplicate
IP should be present. There was now present.
To evaluate the dataset captured we adopted a check list of (Bhuyan et al., 2015)
• Dataset contains real world data, it does but the real world data lacks diversity which is
essentialin teaching the different types of attacks and how they are launched.
• Complete and correct labeling should be performed on the dataset, when labeling the
dataset one should be sure that the packet is intrusive or not. Signature-base was adopted
because with it the packet is for certain intrusive or not. Using Wireshark and no network
architecture led to us missing intrusive action and even the packet we label as intrusive
there is a percentage that they were generated from a network device which was
misconfigured.
• Sufficient trace size: The corpus collected was representing of the TCP/IP conversion on
Sudan networks but not of the different types of attacks that have to be illustrated.
• Featureextraction:feature engineering reduces the amount of data chosen, improves
accuracy by only choosing the fields you need to detect the intrusion you want. One of
our research question was with the limited number of features and Wireshark can we
detect network intrusions, the reason for this is that we didn’t want to overwhelm the
students with too many features and we wanted the dataset size not to increase,.
• Diverse attack scenarios: this was not achieved because the conversion captured did
contain different types of attacks
• Ratio between normal and attack traffic was skewed due to the fact Wireshark drops
packets, thus we used signature base analysis therefore zero base attacks went undetected.
Step One
Capture the data from a live working network
Step Two
Export the packet using csv format

47
Step Three
Run snort to detect Intrusion which will be used in labeling the data
Step Four
Snortgenerates a log , open it in Wiresharkas shown in Wireshark Screen Shot below and then
export it as CSV file
Screenshot of Snort log file opened in Wireshark
Step Five
Export it as CSV file and open in Microsoft Excel . Add a new column name it action label the
data as attack . Add other packets from step two and label them as normal, as illustrated in the
table below.
Table of labeled CSV file opened in Excel Spreadsheet
Step Six
Run several classifiers to choose the best one. Since the data is labeled using supervised learning
,and we choose the classifiers that were easy to explain to our students and were not black boxes.
When we opened the data in Wekawe found there was an imbalance in data as shown in the
screenshotbelow. We first worked with the imbalance data and then used cost sensitive classifier
to try to solve this imbalance.

48
Screenshot demonstrates the imbalance in the data as shown in Weka
Step Seven
Applied decision tree classifier asillustrated in Screenshot below the detection rate was very
high.
Screenshot of Decision tree classifier in Weka

49
Decided to try other classifiers
Naive Bayes
Screenshot of Naive Bayes classifier
Decision table
Screenshot of decision table classifier
The above detection percentage are not logical.
Step Eight
Applied the cost sensitive to balance the data but even after changing the weights the
detection rate was the same.

50
5. CONCLUSION
Collecting the dataset not putting into account where the sensors are located ;before or after the
firewall and not knowing the network map and not creating a baseline of normal behavior makes
it sometimes impossible to differentiate between network related issues and network intrusions.It
has been demonstrated that when apply the network dataset evaluation criteria it was obvious that
the dataset created did not pass the test.Network security prerequisite in networks is a must. It has
been demonstrated without knowledge of networks and protocol analysis it will be hard to detect
intrusions and sometimes impossible. For educational dataset according to our finding it is better
to generate packets using a simulated network environment to represent the different types of
attack classes. The dataset should contain training dataset this should represent different types of
intrusions . If the training dataset is not balance then the result will be skewed.
REFERENCES
[1] Bhuyan, Monowar H., Bhattacharyya, Dhruba K., and Kalita, JugalK.. (2015). Towards Generating
Real-life Datasets for Network Intrusion Detection. International Journal of Network Security,
Vol.17, No.6, PP.675-693.
[2] Ciza, Thomas, Vishwas,Sharma , N. Balakrishan. (2008). Usefulness of Darpra Dataset for Intrusion
Detection system Evaluation. Data Mining, Intrusion Detection, Information Assurance, and Data
Networks Security 2008, 69730G .doi: 10.1117/12.777341
[3] Degadzor, Mohammed, M. A., Effrim, A. F., B. F., &Appiah, K. A. (2017). Brute Force Attack
Detection And Prevention On A Network Using Wireshark Analysis. International Journal of
Engineering Sciences & Research Technology, 6(6), 26-37. doi:10.5281/zenodo.802797
[4] Engels, Andre. (2015). Wikimedia Traffic Analysis Report – OS breakdown per Country. Wikimedia.
9:11. https://stats.wikimedia.org/wikimedia/squids/SquidReportCountryOs.htm
[5] Evans, David. (2009). Introduction to Corpus building and investigation for the humanities, An on-
line information pack about corpus investigation techniques for the Humanities. Available:
https://www.birmingham.ac.uk/Documents/college-artslaw/corpus/Intro/Unit1.pdf
[6] Khan, NiloferShoaib and Lilhore,Umesh. (2016). Review of various intrusion detection methods for
training data sets. International Journal of Modern Trends in Engineering and Research (IJMTER)
Volume 03, Issue 12, ISSN (Online):2349–9745; ISSN (Print):2393-8161
[7] McHugh, John.(2000).The 1998 Lincoln Laboratory IDS Evaluation. RAID 2000, LNCS 1907, pp.
145-161.
[8] Ozg ̈ur,Atilla&Erdem,Hamit.(2016).A Review of KDD99 Dataset Usage in Intrusion Detection and
Machine Learning between 2010 and 2015. Open Access.
https://doi.org/10.7287/peerj.preprints.1954v1
[9] Sangster, Benjamin , O'Connor, T. J. , Cook, Thomas ,Fanelli, Robert , Dean, Erik, Adams, William
J. , Morrell, Chris, Conti, Gregory. (2009). Toward instrumenting network warfare competitions to
generate labeled datasets. Proceedings of the 2nd conference on Cyber security experimentation and
test.p.9-9. Montreal, Canada
[10] Zuech, Richard, Khoshgoftaar, Taghi M., Seliya,Naeem,.Najafabadi, Maryam M, and Kemp,
Clifford.(2015). A New Intrusion Detection Benchmarking System.Florida Artificial Intelligence
Research Society Conference;The Twenty-Eighth International Flairs Conference.Association for the
Advancement of Artificial Intelligence (www.aaai.org)
https://www.aaai.org/ocs/index.php/FLAIRS/FLAIRS15/paper/view/10368/10314

NETWORK INTRUSION DATASETS USED IN NETWORK SECURITY EDUCATION

More Related Content

What's hot (15)

Similar to NETWORK INTRUSION DATASETS USED IN NETWORK SECURITY EDUCATION (20)

Recently uploaded (20)

NETWORK INTRUSION DATASETS USED IN NETWORK SECURITY EDUCATION