Current issues - International Journal of Network Security & Its Applications (IJNSA)

International Journal of Network Security & Its Applications (IJNSA) Vol. 12, No.1, January 2020
DOI: 10.5121/ijnsa.2020.12101 1
A SURVEY ON THE USE OF DATA CLUSTERING FOR
INTRUSION DETECTION SYSTEM IN
CYBERSECURITY
Binita Bohara1
, Jay Bhuyan2
, Fan Wu3
, and Junhua Ding4
1,2,3
Dept.of Computer Science, Tuskegee University, Tuskegee, AL, USA
4
Dept.of Information Science, University of North Texas, Texas, USA
ABSTRACT
In the present world, it is difficult to realize any computing application working on a standalone computing
device without connecting it to the network. A large amount of data is transferred over the network from
one device to another. As networking is expanding, security is becoming a major concern. Therefore, it
has become important to maintain a high level of security to ensure that a safe and secure connection is
established among the devices. An intrusion detection system (IDS) is therefore used to differentiate
between the legitimate and illegitimate activities on the system. There are different techniques are used for
detecting intrusions in the intrusion detection system. This paper presents the different clustering
techniques that have been implemented by different researchers in their relevant articles. This survey was
carried out on 30 papers and it presents what different datasets were used by different researchers and
what evaluation metrics were used to evaluate the performance of IDS. This paper also highlights the pros
and cons of each clustering technique used for IDS, which can be used as a basis for future work.
KEYWORDS
Intrusion detection system, clustering technique, network security
1. INTRODUCTION
Due to the increasing growth of computer network usages, network security is becoming an im-
portant issue. With the evolution of new technology, there is a rapid increase in the incidents of
hacking and intrusion [1]. There is no doubt that all computers are suffering from security
vulnerabilities, that are both technically challenging and costly for manufacturers to fix. So, any
malicious activity on network security vulnerabilities or computers can seriously affect the
system and breach its confidentiality, availability, and integrity. Therefore, the intrusion detection
system has become an integral part of network security architectures.
An intrusion detection system can the ability to locate and identify malicious or anomalous
activities in the computer network by examining the network traffic in real-time [2]. The intrusion
detection system is classified into two groups: misuse and anomaly detection systems. The misuse
intrusion detection system is usually used for commercial purposes because of its predictability
and high accuracy, while anomaly detection is favored in research studies. The intrusion detection
system can also be classified as Host-based and Network-based, depending on the location of de-
tection. There are many number of data mining techniques available for detecting network
intrusions.

2
Data mining is increasingly becoming a popular technique in the network security environment
for finding regularities and irregularities in datasets. Data mining can be defined as the process of
using efficient techniques to extract useful and unexpected patterns from huge datasets [3]. There
are different supervised techniques have been used to detect intrusion. However, this approach
depends on the labeled data and requires the system to be trained on the known data. The problem
with this type of technique is its high dependency on training data and the inability to detect any
new type of attacks. To overcome this limitation of the supervised technique, an unsupervised
approach can be used for intrusion detection, which can detect the unlabeled data. Clustering is
one of the commonly used unsupervised techniques for classifying huge datasets and detecting
the intrusions. Clustering is a data mining technique that is used to infer the conclusion from
unlabeled data and find hidden patterns in datasets [3]. In other words, clustering is a technique
that is used to group similar data into one cluster and other dissimilar data into the different
clusters.
Figure 1: A schematic of clustering process showing three clusters (one red triangle, one black circle, and
one pink plus cluster)
The clustering approach is based on two assumptions [4]. The first assumption is that the number
of normal connections is larger than abnormal connections and the second assumption is that the
feature of the abnormal network is different than the normal network. Depending upon these
assumptions, there are many clustering techniques can be used for detecting intrusions and
attacks such as hierarchical, partitional, grid-based, and density-based clustering techniques.
Hierarchical clustering is a clustering algorithm that constructs a hierarchy of clusters and it
clusters the objects based on their distance. This technique is based on the cluster proximity
measure and there are three measures such as single-linkage, average-linkage, and complete
linkage. Partitional clustering splits the data points into many or some number of separate
partitions, where each partition is known as clusters. K-means is one of the simplest and efficient
partitional clustering algorithms that is used for detecting intrusions in a computer system. Meng
et al [5] used k-means algorithms on KDD99 datasets to detect unknown attacks in different
settings. Density-based clustering algorithm clusters data points based on the density of the points
in a region. DBSCAN is one of the best density-based algorithms. Yang Jian et al [6] used
improved intrusion detection based on DBSCAN that generates clusters depending upon the
density-based method.
This paper presents a literature review on clustering techniques applied in the intrusion detection
system literature. Articles that used clustering technique for intrusion detection were carefully
reviewed and the type of attack different clustering techniques can detect were analyzed. The

3
different clustering techniques were compared in terms of different evaluation metrics used,
datasets used, their strengths and weakness.
The remaining part of the paper is divided into different sections. Section II describes the datasets
used in intrusion detection research. Section III outlines the related definition of the terms used in
this paper. Section III also describes the different terms used in the clustering technique literature.
Section IV lists the related work of the clustering technique used for intrusion detection. Section
V lists the research questions used in this research. Section VI illustrates the results and
discussion. Section VII concludes the paper.
2. DESCRIPTION OF DATASET
KDDCup99 dataset, a publicly available dataset, is the most common dataset that is being used
for evaluating intrusion detection systems [7]. This dataset was used in the KDDCup99
competition and is based on the DARPA98 IDS evaluation dataset [8]. The dataset comprises of 4
GB of compressed TCP dump data of nearly 7 weeks of network traffic. The KDDcup99 dataset
comprises of training and test datasets, where the training dataset consists of nearly 5 million
datasets and the test dataset consists of 3 million datasets [9]. The datasets contain 41 features and
classes: normal and attack. There are 24 different types of attacks in this dataset, which belongs to
four types of major attacks such as:
Denial of Service (DOS): This is an attack where attackers make the system too busy for a
legitimate user to be able to use the system or resources.
User to Root Attack (U2R): U2R starts with the attacker getting access to a normal user account
on a system and exploits its vulnerability so as to get root access to the system.
Remote to Local Attack (R2L): R2L is an attack where an attacker sends packets to the system
over the network and exploits its vulnerability so as to gain access to the system as a normal user.
Probing Attack: This is an attack where an attacker scans the system or network to collect
information so as to identify the potential vulnerability of the system.
Table 1: Different types of Attack from Kddcup 99 Datasets [1]
DOS Probe R2L U2R
Back Nmap Spy Buffer overflow
Land Portsweep Phf Rootkit
Neptune Ipsweep Multihop Loadmodule
Pod Satan ftp write Perl
Smurf Imap
Teardrop Warezmaster
Guess passwd
The new version of KDD99 datasets that is publicly available is called NSL-KDD99 datasets. To
solve the inherent problem of KDD-99, NSL-KDD was proposed. It contains only selected
records from KDD datasets. This new dataset overcame shortcomings of KDD datasets; there
were no redundant records in the training set, which removed the bias in learning algorithms
towards frequent datasets. Similarly, the KDD duplicate data was also removed from the test set.

4
NSL-KDD data consists of 21 types of attacks in the training set and 37 different types of attacks
in the test set. The attacks are classified into four different types: DOS, Probe, U2R, and R2L
[10]. DARPA 98 dataset is also one of the popular datasets used in the intrusion detection system
to evaluate detection rate and false alarm rate in network traffic. This dataset consists of four
different major attack types: DOS, Probing, U2R, and R2L. However, this dataset faced lots of
criticism since a model that was used to generate the traffic was too simple [10].
Both KDD-99 and NSL-KDD do not reflect the real data flow in a computer network as they are
generated on virtual networks on simulation. However, Kyoto 2006+ dataset consists of real
datasets collected over 3 years from November 2006 to August 2009. This dataset comprises of
14 statistical features derived from KDD along with 10 additional features for analysis and
evaluation of intrusion detection system network. These datasets were captured using honeypots,
darknet sensors, email server, and web crawler [10].
Another type of dataset that is used in the literature reviewed in this research for intrusion dataset
is ISCX-2012 intrusion evaluation dataset. This dataset consists of 1512000 packets with 20
features and was obtained by observing network traffic for seven days. There were 75372 normal
traces and 2154 attack traces in training datasets; while in case of testing datasets, there were
19202 normal traces and 37159 attack traces [11].
3. RELATED TERMINOLOGY
• Detection Rate (DR): Detection Rate (DR) can be referred to as a number of attacks
detected by the system divided by the total number of intrusions in the dataset. DR can be
calculated as: DR = True Positive (TP)/(Total number of intrusions in dataset) [12].
• False Alarm Rate (FAR): False Alarm Rate (FAR) is the number of normal instances
classified as attacks divided by the number of normal instances in the dataset. FAR can
be calculated as: FAR= False Positive (FP)/(FP+ True Negative (TN) [12]
• False Negative Rate (FNR): False Negative Rate (FNR) represents the number of
intrusions that were wrongly identified as normal [12]
• True Positive Rate(TPR): True Positive Rate corresponds to IDS that correctly identifies
network activity as a malicious attack [12].
• True Negative Rate(TNR): True Negative Rate occurs when normal instances are
detected as normal [12].
• Accuracy: Accuracy is defined as the percentage of correctly classified instance in
datasets. Accuracy can be calculated as: Accuracy=(TP+TN)/((TP+TN+FP+FN)) [12]
• Sum of Square Error (SSE): The Sum of Square Error (SSE) is the sum of square of
difference between each data point and its cluster mean.
4. RELATED WORK
The intrusion detection system plays an important role in detecting any malicious activities on a
computer network. Various clustering techniques have been designed and implemented for
detecting the intrusions.
Leung et al [13] were able to get a high detection rate while suffering from a highfalse positive
rate using fpMAFIA. The performance value of fpMAFIA was 0.867, which was around 3%-
12% worse off than other algorithms. However, this density and grid-based clustering had the

5
disadvantage of not evaluating each and every point, which made the anomalies clustered into a
set of small clusters and their identification was just more straightforward.
Leonid Portnoy [14] conferred the algorithm that could automatically detect both known and
unknown intrusion. The author used a single-linkage hierarchical clustering method to distinguish
abnormal activities from normal activities. In this algorithm, at the first number of empty clusters
was formed and the distance of instances to the cluster center was checked. The calculated
shortest distance was then compared with the predefined cluster width (W) and if that distance
was shorter than W, then that record was assigned to that cluster. Otherwise, a new cluster was
formed and other instances were assigned to that cluster. The clusters were then updated to the
mean of the cluster centers. The series of updating and reassignment took place until the cluster
center was no more updated. With the algorithm, the average detection rate was around 44- 55%
and the false positive rate was detected 1.3-2.3. Even though this algorithm overcame the
shortcoming of K-means clustering to predict the number of clusters, it still had some limitations.
One of the limitations was in determining the cluster width, which has to be determined
manually. Therefore, there was a chance of mislabeling the normal instance as an abnormal and
vice-versa if W was not determined properly.
Meng et al. [5] used K-means algorithm clustering technique to detect unknown intrusions in a
computer network. This algorithm was used on the KDD cup 1999 dataset, where a number of
clusters were chosen to be 5. Using different settings, the DR with this method was always found
to be above 96 and FAR was below 4 and time complexity was low. This showed that the K-
means algorithm was an effective method for intrusion detection. However, with the K-means
algorithm, cluster number needs to be defined correctly in order to get correct intrusion detection.
In addition, it is also sensitive to the categorical attribute and the results of this algorithm are
usually not steady. In other words, the K-means clustering result can vary, even with the same
input parameter. Similarly, Gerhard et al. [15] also applied a K-means clustering technique for
intrusion detection. The technique was used to classify normal and abnormal network traffic flow
using cluster centroids as a pattern to detect intrusion.
Guan et al. [16] used Y means clustering algorithm to form a number of ‘normal’ and ‘abnormal’
clusters. This algorithm was similar to that of K-means where datasets were partitioned into k
clusters ranging from 1 to n (total number of instances). The second step was to detect the empty
clusters, where new clusters were created to replace empty clusters; followed by re-assigning
instances to existing centers. This process was iterated until there were no empty clusters, which
eventually led to the removal of outliers (splitting) to form new clusters (merging). The last step
in this process was to label the clusters based on their population, meaning that if the population
of the certain cluster was higher than a given threshold of 2.32 ( is the standard deviation of the
data), then the population was re-classified as normal which otherwise was labeled as an
intrusion. H- means+ algorithm was used for clustering KDD-99 dataset with different initial
values of k and showed that the decline of SSE (Sum of Square Errors) was fast when the value
of k was very small. After the turning point (k = 20, DR = 79 % and FAR = 1 %) of k, the
decline of SSE was very slow and the dataset was partitioned into small clusters that were closer
to each other, so there was no decrease in the value of SSE. On average, this method detected
86.63% of intrusion with a false alarm rate of 1.53%. The Y-means algorithm was again trained
with 12,000 unlabeled KDD-99 dataset. The test on the trained system with 10,000 labeled
instances had 82.32% DR and 2.5% FAR indicated that Y-means is one of the more promising
clustering methods for intrusion detection.

6
Zhou et al. [4] proposed a graph-based clustering algorithm as an intrusion detection algorithm to
differentiate between regular and irregular connections. The graph-based algorithm can identify
clusters of any shape and it only uses a parameter and does not require to define any cluster
number. This algorithm used an outlier detection method, which was based on the local deviation
coefficient with different values for k (5,8,11,15). Although there was not much difference
between DR and FPR depending on the value of k, the results showed that DR was higher
(95.3%) and FPR was lowest (2.08%) for k equals to 8.
Wei Jiang et al. [17] used the improved fuzzy clustering algorithm for intrusion detection. The
results showed that their detection rate was much higher with a lower false-positive rate over
KDD’s 99 datasets. In this procedure, the KDD-99 dataset was randomly grouped into 5 groups
with 10 thousand records in each group. The average detection rate was 78.66% and a false
positive rate of 0.704. The proposed method overcomes the disadvantage of fuzzy clustering by
adding weighted value to the data object’s membership degree and optimizes the clustering
number by introducing validity function.
M.Jianliang et al. [5] used a K-means clustering algorithm for intrusion detection to detect the
unknown attack. The algorithm was used in KDD-99 dataset, where the number of clusters (k)
was 5. The experiment was carried out in a different setting and in every setting, the detection
rate was always greater than 96% and a false alarm rate was below 4%. Even though the K-means
clustering technique is used to detect unseen attacks and partition large data it has disadvantages
of degeneracy and cluster dependency.
Li Xue–Yong and Gao Guo [18] proposed improved density-based clustering algorithm IIDBC
for intrusion detection to improve the drawbacks of the density-based clustering algorithm (DB-
SCAN) by using the rational method in merging clusters and calculating the distance. The result
showed that IIDBC improved the performance of DBSCAN, increased the detection rate and de-
creased the false-negative rate. The average detection rate was around 92.33/
Z.Muda et al. [19] used a new intrusion detection algorithm by combining clustering and
classification techniques. This algorithm was a hybrid learning approach based on a combination
of K-means clustering and OneR classification and used KDDcup 99 dataset. The main objectives
of this algorithm were to separate the potential attack from normal instances into different
clusters. This hybrid intrusion detection algorithm has the accuracy and detection rate above 99%
and a false alarm rate below 2.75%. It was found that the performance of the hybrid classifiers
was higher as compared to the single classifier. This algorithm was capable of classifying most of
the instances correctly; however, it could not classify U2R and R2l attack.
Chandrashekhar et al. [20] proposed a new approach based on K-means, fuzzy neural networks
and SVM classifiers as an intrusion detection technique. This technique uses KDDcup99 datasets
to perform the experiment. The proposed technique achieved an accuracy of 97.78% and was
effective for low-frequent attacks such as U2R and R2L.
Ravi Ranjan and G.Sahoo [12] presented a new clustering technique for detecting anomaly
intrusion and attacks. They used the K-medoids method for clustering and the proposed algorithm
achieved high detection rate and overcame the K-means algorithm defects. This approach has
advantages over the existing algorithm such as dependency on initial centroids, cluster number,
and irrelevant clusters. The detection rate for this proposed approach was 91.2% and the false

7
alarm rate is 3.2%. However, the detection rate for the proposed algorithm was low for probing
attack (70.51%) and user to root attack (70.13%).
Chitrakar et al. [21] used a hybrid approach for anomaly-based intrusion detection by combining
Na¨ıve Bayes classification and k-Medoids clustering. This hybrid approach showed better
performance as compared to k-means and Naive Bayes hybrid algorithm. The algorithm used
Kyoto 2006+ dataset and the algorithm provided 4% improvement in terms of accuracy and
detection rate and the false alarm rate was reduced by 1%.
Zhiengje et al. [22] used particle swarm organization with the combination of K-means clustering
method for intrusion detection technique on KDD Cup 1999 data set. The best performance for
this method was when the false detection rate was 2.8% and the detection rate was 82%. The
invasion detected by this technique can be broadly categorized as Probing, DoS, U2R, and R2L.
This method helped to detect known attacks up to 75.82% and unknown attacks to 60.8%. This
method has relatively low detection for DoS, which was due to mislabeling of abnormal data as
normal.
Similarly, Lizhong et al. [23] also used K-means clustering with particle swarm organizations for
anomaly intrusion detection. The algorithm was used in KDD-99 datasets and the experimen- tal
results on PSO-KM showed the detection rate as 86% and false positive rate as 2.8%. The
experiment shows that the accuracy was very good for attack type U2R (78%) and DOS (94%).
However, the accuracy was very low for R2L at 22%.
Fatma et al. [24] used two-stage detection technique in order to improve the detection rate on
DARPA dataset. The detection technique for the first stage was a self-organizing map (SOM)
with K-means algorithm and neural GAS with fuzzy c-means algorithm. The second stage
included SOM with K-means algorithm, support vector machine and decision trees. The results
showed that for the first stage: 69%of the alerts were false attacks and only 31% were real threats,
and for the second stage: 88.77% were false alerts and only 21.23% were real threats. In the first
stage, the neural GAS technique provided better results since the main objective of the first stage
was to cluster low-level alerts into meaningful partitions. In the second stage, the SVM algorithm
provided good result to reduce the rate of false attacks.
Zhong et al. [25] used Kyoto 2006+ and KDD Cup 1999 dataset to evaluate the intrusion
detection by grid-based clustering technique. The results showed that the performance was
insensitive to the variation of the convergence criterion of clustering, attack, and normal
condition in labeling. With the variation on stop condition of cell split, the FPR was 3.25 and DR
was 58.51% for 0.001 value of and FPR and DR value were 5.14% and 62.29%, respectively for
equivalent to 500.
Horng et al. [26] used KDD Cup 1999 training set for detecting different attacks, which were
broadly classified as DoS, U2R, R2L, and Probe. An SVM-based intrusion detection system was
combined with hierarchical clustering algorithm for IDS, which had an accuracy of 95.72 with
only 0.73 false-positive rate. The detection for DoS and Probe were 99.5 and 97.5 respectively,
which were comparatively higher than ESC-IDS [27], KDD’99 winner [28], KDD’99 runner-up
[29], Multi-classifier [30] and Association rule [31].
Lin et al. [32] used KDD-Cup 99 dataset for intrusion detection, where the attacks were classified
as normal, probing, denial of service (DoS), remote to local (R2L), and user to root (U2R). They

8
introduced cluster center and k-nearest neighbor approach known as CANN. In this algorithm,
different evaluation metrics such as accuracy, detection, and false alarms were considered to
evaluate the performance of intrusion detection and they were calculated using a confusion
matrix. The experiment was carried out on 6-dimensional and 19-dimensional KDD dataset [32].
The results showed that the CANN algorithm has an accuracy of 99.76%, detection rate 99.9%
and false alarm rate of about 0.003 for 6- dimensional datasets. Even though the performance was
shown best, CANN totally misclassified U2R and R2L as normal whereas in case of 19-
dimensional KDD dataset, CANN was able to correctly classify U2R and R2L as attacks.
However, the accuracy rate was very low (3.846 and 57.016).
Yongguo et al. [33] proposed a new detection algorithm known as the IDBGC algorithm,
Intrusion Detection Based On Genetic Clustering. The algorithm is a combination of two stages,
nearest neighbor clustering and genetic optimization. The average detection rate for this
algorithm was observed around 60% and the false-positive rate was only 0.4%. The results
showed that the IDBGC algorithm is feasible and effective in detecting unknown intrusions. More
than 50% of DoS, R2L,and U2R unknown attacks were detected. However, for PROBE attack,
the detection rate was less than 50%, which is relatively very low.
Sanjay et al. [34] proposed K-means clustering via na¨ıve bayes classification algorithm for
detecting anomaly-based network intrusion. It was observed that the above algorithm performed
better on KDD cup 99 datasets compared to naive bayes classification in terms of detection rate.
The performance of the algorithms was evaluated based on the accuracy, detection rate, and false
alarm rate. The detection rate was observed as 99% for K-means clustering via na¨ıve bayes
algorithm whereas the false alarm rate was 4. The algorithm was shown to be efficient in
detecting network intrusion; however, this approach had a high false-positive rate.
Amuthan et al. [35] proposed the anomaly network detection method based on K-means
clustering and C4.5 decision tree algorithm. K-means clustering was used first to partition the
training data into k clusters and then decision tree, C4.5 decision tree, was built on each cluster.
The experiment was carried out on KDD99 datasets and the performance was evaluated using
metrics such as true positive rate or detection rate, false-positive rate, precision, accuracy, and F-
measure in percentage. The true positive rate obtained was 99.6%, the false-positive rate was
0.1%, precision was 95.6%, accuracy was 95.8% and F-measure was 94.0. The proposed
algorithm gave a notable detection rate.
Reda et al. [36] proposed a hybrid network intrusion detection system based on random forests
and weighted K-means. The proposed algorithm was evaluated on different percentages of KDD-
99 datasets and the result showed that it achieved high detection rate as 99% and very low false-
positive rate as 12.6 in anomaly detection method whereas in case of misuse detection rate, the
detection rate was 92.73% and relatively very good false positive rate as 0.54%.
Chih-Fong et al. [37] introduced a triangle area based nearest neighbor (TANN) approach to
detect the intrusions in a network. The TANN approach to intrusion detection is a combination of
K- means clustering and k-NN classifier. The approach was used in KDD99 dataset with 10-fold
cross-validation and the result showed that TANN can effectively detect the intrusion detection
with high accuracy and detection rates as compared to support vector machines, k-NN, and the
hybrid classifier K-means and k-NN. The accuracy rate for the proposed algorithm was 99.01%,
detection rate was very high at 99.27, and false-positive rate was 2.99%.

9
Li Tian et al. [38] introduced improved K-means based on the k-medoids cyclic method and the
improved triangle trilateral relations theorem to detect network intrusion detection. This cluster
algorithm for network intrusion detection used KDDcup 99 datasets. The dataset was divided into
5 subsets and the average detection rate was obtained as 89.5%, whereas the false alarm rate was
4.896%, which is relatively very high.
Witcha et al. [39] proposed a fuzzy rough c-means clustering algorithm for detecting new in-
trusions so as to improve the detection rate and reduce false-positive rate in intrusion detection
systems. The fuzzy clustering algorithm was used to predict normal and suspicious behavior. The
performance of the algorithm was measured based on the detection rate, accuracy, false alarm
rate, and correlation. The detection rate for this algorithm was 91.45, accuracy was 82.46, false
alarm rate was 24.8, and correlation was 0.556.
Warusia et al. [11] also integrated K-means clustering and na¨ıve bayes classification
(KMC+NBC) for anomaly intrusion detection for the intrusion detection system. The algorithm
was assessed using ISCX 2012 datasets. The dataset was classified as training and testing datasets
and KMC+NBC algorithm proved to enhance the accuracy and detection rate and reduced false
alarm rate of NBC. The accuracy was 99.8%, the detection rate was 95.4% and the false alarm
rate of the above algorithm was 0.13 in terms of training data; whereas, in testing data, the
accuracy was 99%, the detection rate was 98.8% and false alarm rate was 2.2%.
Vipin Kumar et al. [40] used NSL-KDD datasets to cluster the data into normal and four different
types of attacks i.e., DoS, R2L, U2R, and probe. The experiment was carried out using WEKA
software with clustering technique. Simple K-means algorithm was used to cluster the data into
four different groups of attacks. The algorithm was proven to be very useful in detecting a large
amount of unlabeled data, i.e., detecting new types of attacks.
Abhaya et al. [41] proposed an efficient intrusion detection method for detecting normal and
abnormal instances based on the fuzzy c-means clustering technique and support vector machine.
They used KDD99 datasets for this approach and compared the performance with K-means,
SVM, K-means and na¨ıve bayes, and FCM and na¨ıve bayes. In this approach, the datasets are
transferred to the clustering technique and it is transferred to the classification technique and
finally, the performance is evaluated using accuracy. The accuracy of FCM+SVM was 99.74% as
compared to NB+FCM, SVM+k-means, NB+k-means which accounts to 95.63%, 99.37% and
95.32%.
Hari Om and Aritra Kundu [42] proposed hybrid intrusion detection by combining K-means and
two classifiers, na¨ıve bayes and k-nearest neighbor, to detect anomaly. This algorithm used
KDD-cup99 datasets to detect the intrusion and classify them into four categories: DoS, U2R,R2l,
and probe. The main objective of the proposed algorithm was to reduce the false alarm rate. The
detection rate of the proposed approach was 99.35%, the false alarm rate was 1.394%, and
accuracy was 98.20%.
Partha et al. [43] evaluated the data mining technique based on the fuzzy c-means clustering and
K-means clustering technique over the NSL-KDD datasets to detect the four types of attacks. It
was observed that only 45.95% of attacks were detected by fuzzy c-means clustering technique,
whereas the detection rate is 44.72% using K-means clustering technique.

10
5. RESEARCH QUESTIONS
The main objective of the research is to analyze the clustering techniques used in the literature of
the intrusion detection system over 10 years. To achieve this objective, the following questions
were established:
1. What datasets have been used in IDS?
2. What clustering technique has been used in the intrusion detection system research?
3. What are the evaluation metrics used to measure the performance of clustering technique?
4. What are the strengths and weakness of clustering techniques that have been used in IDS?
6. RESULT AND DISCUSSION
The result for this research was analyzed using the research question that was initially posted.
There were four questions that were to be answered, which we analyzed separately as following:
1. RQ1: What was the dataset that has been used in intrusion detection system?
There are different datasets being used by different learning algorithm such as classification,
clustering, regression, etc. to detect the intrusions. In this survey, 30 papers were reviewed for
clustering technique. The most common type of dataset that is being used for clustering technique
is KDDcup99 datasets. The number of papers using KDD-99 dataset had peaked in year 2011.
Among 30 papers, there were 24 that used KDD-99 dataset, where Zhong et al. used KDD99 data
along with Kyoto 2006+ datasets accounting to 80% of whole datasets. KDD- 99 dataset is
derived from DARPA dataset, this is one of the datasets that is used by researchers for
clustering algorithm. In addition, only 2 papers used NSL-KDD dataset and 2 papers used Kyoto
2006+ that account for 7% of the datasets used in this research as stated in figure 2. Similarly,
only one paper each reviewed in this research used DARPA 98 datasets and ISCX 2012 datasets.
Figure 2: A schematic view of datasets used in clustering technique

11
Figure 3: A schematic view of distribution of datasets over the years
2). Q2: What clustering technique has been used in the intrusion detection system research?
One of the major questions about this research is to find out what is the most common clustering
technique used in intrusion detection system research. Out of 30 papers studied, the most
common clustering technique used for intrusion detection was K-means (it was used 14 times).
This clustering technique was used as a single technique for three times while for others, it was
used as a hybrid technique such as hierarchial, One each R, fuzzy and SVM, Na¨ıve Bayes, PSO,
fuzzy C-means, SOM and neutral gas, and C 4.5 decision tree. Besides these graph-based, II
DBC, K-medoids, grid-based, SUM + hierarchical, CAAN, IDBGC, Random forest + weighted,
and TANN were used.
3). RQ3: What are the evaluation metrics used to measure the performance of the clustering
technique?
The clustering technique used in intrusion detection for computer networks can be assessed and
evaluated using various types of evaluation metrics. The detection rate, false alarm rate, false-
positive rate, false-negative rate were the most common evaluation metrics used to measure the
performance of clustering technique in the clustering detection literature. All the 30 papers
reviewed used these metrics for detecting the performance of the clustering technique in the
intrusion detection system. However, Leung et al. [13] and Zhiengje et al. [22] also used ROC
curve while evaluating the techniques; and similarly, Guan et al. [16] also considered sum of
square error along with detection rate and false-positive rate. Chan- drashekar et al. [20] also
looked at specificity and sensitivity for the datasets in addition to TPR, FPR, DR, FPR and FNR.
The performance of the clustering technique is considered high if the detection rate, accuracy is

12
high and the false alarm rate is low. These evaluation matrices were also used to compare the
algorithms with the already existed algorithm.
Table 2: Strengths and Weakness of Clustering Technique used in intrusion detection literature
Clustering
Technique
Strength Weakness Ref
K-means DR was high and FAR
was below 4% and time
complexity was low
Cluster number needs to
be defined
Meng et al.,
Gerhard et al.,
Vipin et al
Hierarchical
clustering
overcame the
short-
coming of
K-means
clustering to
predict the
number of
clusters
Determining the cluster
width manually, there was a
chance of mislabeling the
normal instance as an
abnormal and vice-versa if W
was not determined properly
Leonid et al.
Y-means Number of cluster depen-
dency and dengenracy of K-
means was overcome
High false positive com-
pared to other algorithm
Guan et al.
Graph-based Identifies clusters of any
shape and it only uses a
parameter and does not require
to define anyclus- ter number
Computation increased as
number of records increased
Zhou et al
k-medoids Has advantages over the
existing algorithm such as
dependency on initial
centroids, cluster number, and
irrelevant clusters
The detection rate for
the proposed algorithm was
low for probing attack(70.51)
and user to root attack(70.13)
Ravi et al.
IFCA Introduce the
function
of validity for choosing
number of clustering
Wei et al.
IIDBC Detection rate was high
and the performance of
DBSCAN was improved
Selection of Parameters Li-XUE et
al.
Grid-based The performance was in-
sensitive to the variation of the
convergence criterion of
clustering, attack and normal
condition
Low detection rate and
high false positive
Zhong et al.
CANN Detection rate was high
for 6-dimensional KDD
datasets
Misclassified U2R and
R2L as normal in case of 6-
dimensional KDD datasets
Lin et al.
TANN The accuracy rate, detec-
tion rate were high
Chih-Fong
et al.

13
Improved K-
means
False alarm rate was rela-
tively very high
Li Tian et al.
Fuzzy
C-means
Achieved good per-
formance compared
to Kmeans methods
Witcha et al
Density+Grid
based
High detection rate High false positive rate Leung et al.
K-
means+One R
classification
Detection rate above 99.0
and a false alarm rate below
2.75 and the performance of
the hybrid classifiers was
higher as compared to the
single classifier
Couldn’tclassify U2R
and R2l attack
Z.muda et al.
Kmedoids+
Naive Bayes
Showed better perfor-
mance as compared to K-
means and Naive Bayes
hybrid algorithm
Chitrakar et
al.
PSO+K-
means
Accuracy was very good
for U2R and DoS
Low detection rate and
high false positive
Lizhong et
al., Zhiengje et
al.
SVM+HC Detection rate was high
for all four types of attack
Horng et al.
K-
means+Naive
Bayes
The algorithm was shown
to be efficient in detecting
network intrusion
This approach had a high
false-positive rate
Sanjay etal
Warsula et al.
K-
means+C4.5
The proposed algorithm
gave notable detection
rate
Amuthan et
al.
Random
For-
est+Weighted
K-means
High false-positive rate Reda et al.
Fuzzy
Cmeans+Svm
The accuracy value was
slightly better than existing
K-Medoids+SVM and it was
far better than SVM alone and
it was shown to be stable over
other
Abhaya et
al.
SOM+
K-means+
Fuzzy Cmean
Showed to outperform
other competitor method by
reducing FAR
Fatma et al
Kmeans +
Fuzzy + SVM
Effective for low-frequent
attacks such as U2R and R2L
Chandrashekar
et al.
K-means +
KNN + NB
May sometime misclas-
sify the records
Hari Om et
al.
Fuzzy
Cmeans+K-
means
Very low detection rate Partha et al.

14
4) RQ4) What are the strengths and weakness of clustering techniques that have been used
in IDS?
After reviewing 30 papers in this research, we also evaluated the strengths and weakness of the
clustering techniques used. The common techniques used by different research were combined
and analyzed for strengths and weakness. The results were then incorporated as shown in table 2.
From table 2, it can seen that each and every clustering technique has its own pros and cons in
analyzing intrusion detection.
7. CONCLUSION
The intrusion detection system is a way of analyzing the network traffic so that unwanted packets
or any malicious activities on the system are detected and prevented. In this paper, 30 papers of
clustering techniques were studied and evaluated for intrusion detection. It was found that 27
different types of clustering techniques were applied in order to detect intrusions. The most
common clustering technique used for intrusion among the 30 papers was K-means clustering
technique.
Apart from using a single clustering technique, there was a hybrid clustering technique that was
combined with different classification techniques for detecting intrusions in a system. These
clustering techniques were used in different datasets. However the most common datasets were
found to be KDD-99 datasets. The KDDCup99 datasets comprise of millions of connections with
different features and different types of attacks. Similarly, the performance of this technique was
evaluated by different researchers using different evaluation metrics. Detection rate and False
Alarm rate was commonly used as evaluation metrics by most of the researchers to check the
effectiveness of their procedures for intrusion detection. Similarly, in this research, the strength
and weakness of different technologies used for clustering attacks were also analyzed.
As future work, we are researching on the effectiveness of user-oriented clustering to enhance the
effectiveness of an Intrusion Detection System. In this method, clusters of activities in the log
files are formed based on an expert feedback of whether or not the activities are an intrusion. The
cluster is defined based on the activities it contains. A future activity is defined as intrusion or not
based on its similarity to a cluster.
8. ACKNOWLEDGEMENTS
This research was supported in part by National Science Foundation grants #1761735, #1723586,
#1663350. National Institutes of Health grant NIH TU CBR/RCMI #U54MD007585, and a grant
from Rockwell Collins, USA
REFERENCES
[1] S. Revathi and A. Malathi, “A detailed analysis on nsl-kdd dataset using various machine learning
techniques for intrusion detection,” International Journal of Engineering Research & Technology
(IJERT), vol. 2, no. 12, pp. 1848–1853, 2013.
[2] F. Salo, M. Injadat, A. B. Nassif, A. Shami, and A. Essex, “Data mining techniques in intrusion
detection systems: A systematic literature review,” IEEE Access, vol. 6, pp. 56 046– 56 058, 2018.

15
[3] M. K. Siddiqui and S. Naahid, “Analysis of kdd cup 99 dataset using clustering based data mining,”
International Journal of Database Theory and Application, vol. 6, no. 5, pp. 23–34, 2013.
[4] Z. Mingqiang, H. Hui, and W. Qian, “A graph-based clustering algorithm for anomaly intrusion
detection,” in 2012 7th International Conference on Computer Science & Education (ICCSE).
IEEE, 2012, pp. 1311–1314.
[5] M. Jianliang, S. Haikun, and B. Ling, “The application on intrusion detection based on k-means
cluster algorithm,” in 2009 International Forum on Information Technology and Applications, vol. 1.
IEEE, 2009, pp. 150–152.
[6] Y. Jian, “An improved intrusion detection algorithm based on dbscan [j],” Microcomputer
Information, vol. 25, no. 13, 2009.
[7] K. Cup, “Dataset,” available at the following website http://kdd. ics. uci.
edu/databases/kddcup99/kddcup99. html, vol. 72, 1999.
[8] R. P. Lippmann, D. J. Fried, I. Graf, J. W. Haines, K. R. Kendall, D. McClung, D. Weber,
S. E. Webster, D. Wyschogrod, R. K. Cunningham et al., “Evaluating intrusion detection systems:
The 1998 darpa off-line intrusion detection evaluation,” in Proceedings DARPA Information
Survivability Conference and Exposition. DISCEX’00, vol. 2. IEEE, 2000, pp. 12–26.
[9] M. Tavallaee, E. Bagheri, W. Lu, and A. A. Ghorbani, “A detailed analysis of the kdd cup 99 data
set,” in 2009 IEEE Symposium on Computational Intelligence for Security and Defense Applications.
IEEE, 2009, pp. 1–6.
[10] D. D. Proti c´, “Review of kdd cup’99, nsl-kdd and kyoto 2006+ datasets,” Vojnotehnicˇki
glasnik, vol. 66, no. 3, pp. 580–596, 2018.
[11] W. Yassin, N. I. Udzir, Z. Muda, M. N. Sulaiman et al., “Anomaly-based intrusion detection through
k-means clustering and naives bayes classification,” in Proc. 4th Int. Conf. Comput. Informatics,
ICOCI, no. 49, 2013, pp. 298–303.
[12] R. Ranjan and G. Sahoo, “A new clustering approach for anomaly intrusion detection,” arXiv preprint
arXiv:1404.2772, 2014.
[13] K. Leung and C. Leckie, “Unsupervised anomaly detection in network intrusion detection using
clusters,” in Proceedings of the Twenty-eighth Australasian conference on Computer Science-Volume
38.Australian Computer Society, Inc., 2005, pp. 333–342.
[14] L. Portnoy, “Intrusion detection with unlabeled data using clustering,” Ph.D. dissertation, Columbia
University, 2000.
[15] G. M u¨nz, S. Li, and G. Carle, “Traffic anomaly detection using k-means clustering,” in
GI/ITG Workshop MMBnet, 2007, pp. 13–14.
[16] Y. Guan, A. A. Ghorbani, and N. Belacel, “Y-means: A clustering method for intrusion detection,” in
CCECE 2003-Canadian Conference on Electrical and Computer Engineering. Toward a Caring and
Humane Technology (Cat. No. 03CH37436), vol. 2. IEEE, 2003, pp. 1083–1086.
[17] W. Jiang, M. Yao, and J. Yan, “Intrusion detection based on improved fuzzy c-means algorithm,” in
2008 International Symposium on Information Science and Engineering, vol. 2. IEEE, 2008, pp. 326–
329.

16
[18] L. Xue-Yong, G. Guo-hong, and S. Jia-Xia, “A new intrusion detection method based on improved
dbscan,” in 2010 WASE International Conference on Information Engineering, vol. 2. IEEE,
2010, pp. 117–120.
[19] Z. Muda, W. Yassin, M. N. Sulaiman, and N. I. Udzir, “Intrusion detection based on k-means
clustering and oner classification,” in 2011 7th International Conference on Information Assurance
and Security (IAS). IEEE, 2011, pp. 192–197.
[20] A. Chandrasekhar and K. Raghuveer, “Intrusion detection technique by using k-means, fuzzy neural
network and svm classifiers,” in 2013 International Conference on Computer Communication and
Informatics.IEEE, 2013, pp. 1–7.
[21] R. Chitrakar and C. Huang, “Anomaly based intrusion detection using hybrid learning approach of
combining k-medoids clustering and naive bayes classification,” in 2012 8th In- ternational
Conference on Wireless Communications, Networking and Mobile Computing. IEEE, 2012, pp. 1–5.
[22] Z. Li, Y. Li, and L. Xu, “Anomaly intrusion detection method based on k-means clustering algorithm
with particle swarm optimization,” in 2011 International Conference of Informa- tion Technology,
Computer Engineering and Management Sciences, vol. 2. IEEE, 2011, pp. 157–161.
[23] L. Xiao, Z. Shao, and G. Liu, “K-means algorithm based on particle swarm optimization algorithm
for anomaly intrusion detection,” in 2006 6th World Congress on Intelligent Control and Automation,
vol. 2. IEEE, 2006, pp. 5854–5858.
[24] H. Fatma and L. Mohamed, “A two-stage technique to improve intrusion detection systems based on
data mining algorithms,” in 2013 5th International Conference on Modeling, Sim- ulation and
Applied Optimization (ICMSAO). IEEE, 2013, pp. 1–6.
[25] Y. Zhong, H. Yamaki, and H. Takakura, “A grid-based clustering for low-overhead anomaly intrusion
detection,” in 2011 5th International Conference on Network and System Security. IEEE, 2011, pp.
17–24.
[26] S.-J. Horng, M.-Y. Su, Y.-H. Chen, T.-W. Kao, R.-J. Chen, J.-L. Lai, and C. D. Perkasa, “A novel
intrusion detection system based on hierarchical clustering and support vector machines,” Expert
systems with Applications, vol. 38, no. 1, pp. 306–313, 2011.
[27] A. N. Toosi and M. Kahani, “A new approach to intrusion detection based on an evolutionary soft
computing model using neuro-fuzzy classifiers,” Computer communications, vol. 30, no. 10, pp.
2201–2212, 2007.
[28] B. Pfahringer, “Winning the kdd99 classification cup: bagged boosting,” SIGKDD explo- rations, vol.
1, no. 2, pp. 65–66, 2000.
[29] I. Levin, “Kdd-99 classifier learning contest: Llsoft’s results overview,” SIGKDD explo- rations, vol.
1, no. 2, pp. 67–75, 2000.
[30] M. Sabhnani and G. Serpen, “Application of machine learning algorithms to kdd intrusion detection
dataset within misuse detection context.” in MLMTA, 2003, pp. 209–215.
[31] W. Xuren, H. Famei, and X. Rongsheng, “Modeling intrusion detection system by discov- ering
association rule in rough set theory framework,” in 2006 International Conference on Computational
Inteligence for Modelling Control and Automation and International Con- ference on Intelligent
Agents Web Technologies and International Commerce (CIMCA’06). IEEE, 2006, pp. 24–24.

17
[32] W.-C. Lin, S.-W. Ke, and C.-F. Tsai, “Cann: An intrusion detection system based on combining
cluster centers and nearest neighbors,” Knowledge-based systems, vol. 78, pp. 13–21, 2015.
[33] Y. Liu, K. Chen, X. Liao, and W. Zhang, “A genetic clustering method for intrusion detection,”
Pattern Recognition, vol. 37, no. 5, pp. 927–942, 2004.
[34] S. K. Sharma, P. Pandey, S. K. Tiwari, and M. S. Sisodia, “An improved network intrusion
detection technique based on k-means clustering via na¨ıve bayes classification,” in IEEE-
International Conference On Advances In Engineering, Science And Management (ICAESM-2012).
IEEE, 2012, pp. 417–422.
[35] A. P. Muniyandi, R. Rajeswari, and R. Rajaram, “Network anomaly detection by cascading k-means
clustering and c4. 5 decision tree algorithm,” Procedia Engineering, vol. 30, pp. 174–182, 2012.
[36] R. M. Elbasiony, E. A. Sallam, T. E. Eltobely, and M. M. Fahmy, “A hybrid network intrusion
detection framework based on random forests and weighted k-means,” Ain Shams Engineering
Journal, vol. 4, no. 4, pp. 753–762, 2013.
[37] C.-F. Tsai and C.-Y. Lin, “A triangle area based nearest neighbors approach to intrusion detection,”
Pattern recognition, vol. 43, no. 1, pp. 222–229, 2010.
[38] L. Tian and W. Jianwen, “Research on network intrusion detection system based on improved k-
means clustering algorithm,” in 2009 International Forum on Computer Science- Technology and
Applications, vol. 1. IEEE, 2009, pp. 76–79.
[39] W. Chimphlee, A. H. Abdullah, M. N. M. Sap, S. Srinoy, and S. Chimphlee, “Anomaly- based
intrusion detection using fuzzy rough clustering,” in 2006 International Conference on Hybrid
Information Technology, vol. 1. IEEE, 2006, pp. 329–334.
[40] V. Kumar, H. Chauhan, and D. Panwar, “K-means clustering approach to analyze nsl- kdd
intrusion detection dataset,” International Journal of Soft Computing and Engineering (IJSCE), 2013.
[41] K. Kumar et al., “An efficient network intrusion detection system based on fuzzy c-means and
support vector machine,” in 2016 International Conference on Computer, Electrical &
Communication Engineering (ICCECE). IEEE, 2016, pp. 1–6.
[42] H. Om and A. Kundu, “A hybrid system for reducing the false alarm rate of anomaly intrusion
detection system,” in 2012 1st International Conference on Recent Advances in Infor- mation
Technology (RAIT). IEEE, 2012, pp. 131–136.
[43] P. S. Bhattacharjee, A. K. M. Fujail, and S. A. Begum, “A comparison of intrusion detection by k-
means and fuzzy c-means clustering algorithm over the nsl-kdd dataset,” in 2017 IEEE International
Conference on Computational Intelligence and Computing Research (ICCIC). IEEE, 2017, pp. 1–6.

18
AUTHORS
Binita Bohara is a graduate student of Information System and Security Management at
Tuskegee University. She completed her Accounting degree from Australia. She is
currently working on a research on clustering techniques for intrusion detection for net-
work security under professor Dr. Jay Bhuyan. Currently, she is also working as
a teaching assistant for undergraduate students.
Jay Bhuyan is a professor in the Department of Computer Sci- ence at Tuskegee
University. His research and teaching interests include Telecom Software Architecture
and Development, Software and Network Security, Big Data Analytics, and Machine
Learn- ing. He received a Ph.D. in Computer Science from the Uni- versity of
Louisiana at Lafayette. He has over 25 years of full- time and part-time teaching
experience as well as over 15 years of research and development experience in the
Telecom industry. He is a member of ACM, IEEE and IEEE Computer, IEEE Communications Societies.
Fan Wu is a professor of Computer Science and the director of Tuskegee University
Center of Academic Excellence in Information Assurance Education. He received his
Ph.D. degree in Computer Science from Worcester Polytechnic Institute (WPI) in 2008.
His teaching and research interests lie primarily in the area of Information Assurance,
Soft- ware Security, Mobile Security, High Performance Computing and Ar- tificial
Intelligence. He has led several cybersecurity related projects funded by NSF, DHS, and
DoD. He has served as the editor-in-chief of the International Journal of Mobile Devices,
Wearable Technology, and Flexible Elec- tronics (IJMDWTFE). He has involved students in research and
published a number of research papers in cybersecurity, software security, mobile security, and
information assurance.
Junhua Ding is a Professor of Data Science and the Director of Data Science program in
the department of Information Science at the Univer- sity of North Texas. His research
interests include: data analytics, machine learning, software security and software
engineering. He received a Ph.D., an M.S., and a B.S. in Computer Science from Florida
Inter- national University, Nanjing University, and China University of Geo- sciences,
respectively.

Current issues - International Journal of Network Security & Its Applications (IJNSA)

More Related Content

What's hot (18)

Similar to Current issues - International Journal of Network Security & Its Applications (IJNSA) (20)

Recently uploaded (20)

Current issues - International Journal of Network Security & Its Applications (IJNSA)