Managing Intrusion Detection Alerts Using Support Vector Machines

Majid Ghonji Feshki, Omid Sojoodi Shijani & Minoo Deljavan Anvary
International Journal of Computer Science and Security (IJCSS), Volume (9) : Issue (5) : 2015 266
Managing Intrusion Detection Alerts Using Support Vector
Machines
Majid Ghonji Feshki ghonji.majid@yahoo.com
Department of computer science
Islamic Azad University, Qazvin Branch
Qazvin, Qazvin, Iran
Omid Sojoodi Shijani o_sojoodi@qiau.ac.ir
Department of computer science
Islamic Azad University, Qazvin Branch
Qazvin, Qazvin, Iran
Minoo Deljavan Anvary Minoo.deljavan@yahoo.com
IT Department School of e-Learning
Shiraz University
Shiraz, Fars, Iran
Abstract
In the computer network world Intrusion detection systems (IDS) are used to identify attacks
against computer systems. They produce security alerts when an attack is done by an intruder.
Since IDSs generate high amount of security alerts, analyzing them are time consuming and error
prone. To solve this problem IDS alert management techniques are introduced. They manage
generated alerts and handle true positive and false positive alerts. In this paper a new alert
management system is presented. It uses support vector machine (SVM) as a core component of
the system that classify generated alerts. The proposed algorithm achieves high accurate result
in false positives reduction and identifying type of true positives. Because of low classification
time per each alert, the system also could be used in active alert management systems.
Keywords: Intrusion Detection System, Alert Management, Support Vector Machine, Security
Alert Classification, Reduction of False Positive Alerts, Classifying True Positive Alert Based on
Their Attack types.
1. INTRODUCTION
An intrusion detection system (IDS) monitors network traffic and suspicious activity and then
alerts the system or network administrator. In some cases the IDS may also responds to
anomalous or malicious traffic by taking action such as blocking the user or source IP address
from accessing the network [1]. Common problem with IDSs is producing many alerts each day
that many of them are false positive alerts. Since amount of fake alerts is high so alerts of real
attacks are hided among them. IDSs divided in to two categories, passive and active. 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 as active IDSs, work in online mode. Some of problems
of IDS are: huge amount of generated alerts and high rate of false positive alerts among
generated alerts. Also most alert management system has low speed, and has low accuracy in
classification results. In this paper a new alert management system proposed that uses Support
Vector Machine (SVM) as a classification engine [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 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
Clustering and classification of alerts is a technique of alert management systems. A method of
clustering based on root causes is proposed by K. Julisch [3] 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 technique. In [4, 5] two genetic clustering algorithm based,
named Genetic Algorithm (GA) and Immune based Genetic Algorithm used to manage IDS alerts.
Their proposed methods are depended on underlying network information same as method
presented by Julisch.
Three algorithms with dimension reduction techniques are used to cluster generated IDS alerts
from DARPA 2000 dataset [6]in [4] and then compared results. The problems of that system are:
row alert without preprocessing are entered to the algorithms and system is not tuned. Cuppens
proposed another method as a part of MIRADOR project that uses expert system to make
decision [7, 8].
Debar et al. [9] designed a system by placing them in situations aggregates alerts together.
Situations are set of special alerts. Some attributes of alert are used to construct a situation.
In [10] Azimi et.al. is proposed a new system that manage alerts generated from DARPA 98
dataset with snort. Some algorithm such as alert filtering, alert preprocessing and cluster merging
are used in that system. The main unit of the system is cluster/classify unit that uses Self-
Organizing Maps (SOM) [11] to cluster and classify IDS alerts. Results of [10] show that SOM
was able to cluster and classify true positive and false positive alerts more accurate than other
techniques.
In another work, authors of previous article develop an alert management system [12] similar to
[10]. In that work usage of seven genetic clustering algorithms named Genetic Algorithm (GA)
[13], Genetic K-means Algorithm (GKA) [14], Improved Genetic Algorithm (IGA) [15], Fast
Genetic K-means Algorithm (FGKA) [16], Genetic Fuzzy C-means Algorithm (GFCMA) [17],
Genetic Possibilistic C-Means Algorithm (GPCMA) [12] and Genetic Fuzzy Possibilistic C-Means
Algorithm (GFPCMA) [12] 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 [12].
Also they were develop another approach based on same alert management system in [10, 12]
by replacing classification engine of previous work by Learning Vector Quantization (LVQ) [18,
19]. In that work LVQ technique is used in the classification unit. The experimental results shows
that LVQ has can be a solution to alert management systems that used in active mode because
of high classification speed. The accuracy of the suggested approach is acceptable [18].
In [20] a fast and efficient vulnerability based approach that addresses the above issues is
proposed. It combines several known techniques in a comprehensive alert management
framework in order to offer a novel solution. Their approach is effective and yields superior results
in terms of improving the quality of alerts.
Alert Correlation is one of alert management techniques is used to find any relation between
alerts and tries to find attack scenarios. In a new alert correlation technique named ONTIDS is

proposed. It is a context-aware and ontology-based alert correlation framework that uses
ontologies to represent and store the alerts information, alerts context, vulnerability information,
and the attack scenarios. ONTIDS employs simple ontology logic rules written in Semantic
Query-enhance Web Rule Language (SQWRL) to correlate and filter out non-relevant alerts. they
illustrate the potential usefulness and the flexibility of ONTIDS by employing its reference
implementation on two separate case studies, inspired from the DARPA 2000 [21] evaluation
datasets.
In this paper an alert management system based on system proposed by Azimi et.al. [10] is
proposed that uses SVM as a tool to classify input alert vectors. It covers the problems of
previous work such as low accuracy of results, inability to identify the types of generated alerts
accurately. The system will be able to improve accuracy of results, to detect type of true positive
alerts and also to reduce the number of false positive alerts.
3. USING SVM IN ALERT MANAMENT SYSTEM
In this section we investigate alert management framework introduced by [10, 12]. The suggested
framework consists of several units. In this investigation we use Snort [11] IDS to generate alerts
from DARPA 98 dataset [22]. Figure 1 shows the proposed framework. Snort is an open source,
network based and signature based IDS. Snort gets tcpdump binary files of DARPA 98 dataset as
input and produces security alerts. After producing alerts with snort; they are entered to labeling
unit. The labeling unit and others units are investigated in next sections.
FIGURE 1: Proposed alert management system.
3.1 Labeling Unit
Labeling unit according to tcpdum.list files label generated security alerts. It means that this unit
appends attack type of each alert as an attribute of processed alert to proper alert. This operation
used in next unit to train and to specify correctness of generated results of the system. Figure 2
shows the labeling algorithm.

FIGURE 2: Alert labeling algorithm used in Labeling Unit [10, 12].
3.2 Normalization and Filtering Unit
According to [23] snort is unable to detect all attacks in DARPA 98 dataset then we use “alert
types” file to specify accepted attack types in this investigation [10, 12]. In this unit alerts
according to its labels are filtered. All of accepted attack types are entered to this unit and then
this unit removes all of alerts that their labels are not in “alert types” file. Also this unit removes
redundant alerts from set and keeps one of them. In this paper accepted alerts attack types are:
Back, Land, Pod, Phf, Rootkit, Nmap, Imap, and Dict.
3.3 Preprocessing Unit
This unit transforms character form values of attributes of alert to numerical data to construct data
vectors from alerts (1) and (2). The range of attribute values is reduced by this unit (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 SVM Training and Classification Unit
In this unit SVM algorithm is used as a classifier. SVM is a technique to solve classification
problems [2]. It maps pattern vectors to a high dimensional feature space to generate best
separating hyperplane. SVM uses linear model to implement nonlinear models. When a linear
model constructed in new space can represent nonlinear decision boundaries in original space.
One of main characteristics of SVM is that it simultaneously minimizes the empirical classification
error and maximizes the geometric margin [24]. There are many types of kernels that may be
used in an SVM [25]. Some nonlinear kernel type is Polynomial and Gaussian (radial based
Kernel). Equations Polynomial and Gaussian are respectively.
1. Input TCPDUMP list files.
2. Input alert log files.
3. Create an empty AttackList set.
4. Create an empty AlertList set.
5. For each row in TCPDUMP list files:
5.1. If the row is a labeled attack then add
the row to the AttackList set.
6. For each row in alert log files:
6.1. Create key with the five attributes:
source IP, destination IP, source port,
destination port, ICMP code/type.
6.2. If the key exists in AttackList set then
label the selected row with the type of
found attack from AttackList set.
Else
Label the selected row with the False
Positive attack type.
6.3. Add the selected row to the AlertList
set.
7. Return the AlertList set.

( )dT
yxcoeff + (4)
( )22
/exp δyx −−
(5)
In this paper we use Polynomial and Gaussian kernel functions to evaluate SVM classifier.
4. EXPERIMENTAL RESULTS
To simulate the proposed system C#.net programming language, MATLAB and OSU SVM
toolbox are used[26, 27]. The parameters of simulation are shown below.
In the investigation three type of kernel used for SVM named Linear, Polynomial and Radial Basis
Function (RBF) kernels. Each of these kernels has own parameters for example Linear SVM
depends on C and Epsilon parameters where Epsilon is used to termination threshold. The C
parameter tells the SVM optimization how much you want to avoid misclassifying each training
example. For large values of C, the optimization will choose a smaller-margin hyperplane if that
hyperplane does a better job of getting all the training points classified correctly. Conversely, a
very small value of C will cause the optimizer to look for a larger-margin separating hyperplane,
even if that hyperplane misclassifies more points. For very tiny values of C, you should get
misclassified examples, often even if your training data is linearly separable. Polynomial kernel
depends of four parameters named Degree, Gamma, Coefficient and C. Also RBF kernel
depends on two factor named Gama and C.
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.
To evaluate the performance of algorithms four measurements are introduced, they are:
1- Classification Error (ClaE),
2- Classification Accuracy percent (ClaAR),
3- Average Alert Classification Time (AACT),
4- False Positive Reduction Rate (FPRR).
According to tables 1 to 3 the classification error rate reduces by increasing the value of C.
Results show that RBF kernel is best to classify IDS alerts. According to table 3 when Gama is
10 and C is 16 best results is achieved.
The best values of ClaAR are 99.82%. The value of AACT measurement is 0.0.00032 that shows
the proposed system can be used in active IDS alert management systems that evaluate alerts
beside alert production by IDS concurrently.
The tables show the results of accuracy of proposed system based on Linear, Non-linear and
RBF kernels in identifying attack type of each alert vector in test phase. As it can be seen in table
3, the proposed system can identify all of attack types of alerts with high rate of accuracy when
RBF kernel is used.
An important point is accuracy percent of false positive identification. That is the proposed system
can reduce false positive alerts with 99.94 percent. Which shows to be a solution of an important
problem of IDSs. Proposed alert management system reaches 100 percent for Back, Land, Dict
and Nmap attack types. For attack types Phf, Imap and Rootkit accuracy percent values are
66.67, 66.67 and 42.86 respectively. The number of generated SVM is 211.

Bahrbegi et. al. in [12] proposed a framework that uses genetic algorithm families to clustering
and classification propose. As two works are similar we have to compare our results with their
work. These results are shown in table 3. For all metrics the proposed system has high value in
contrast of all GA based techniques. As shown in table 4, these algorithms could not be able to
work actively because of the execution times are high. Although the proposed method earns high
accuracy results per alert attack type.
C CAR Time SVM
No
Back Land Pod Phf Rootkit Imap Dict NMap FPRR
1 99.01 0.000065 1029 99.22 0 97.96 0 0 0 100 100 99.15
TABLE 1: Extracted performance metric values from simulation with Linear Kernel.
C CAR Time degre
e
SVM No Back Lan
d
Pod Phf Rootk
it
Imap Dict NMap FPRR
1 99.01 0.0002
22
1 1029 99.22 0 97.96 0 0 0 100 100 99.16
1 99.61 0.0000
82
5 337 99.92 0 97.96 33.33 0 66.67 100 100 99.94
1 99.77 0.0000
54
10 238 100 100 97.96 66.67 14.29 66.67 100 100 99.94
0.
5
99.74 0.0005
4
10 209 100 100 97.96 66.67 28.57 33.33 100 100 99.89
1 99.77 0.0000
47
15 209 100 100 97.96 66.67 28.57 66.67 100 100 99.89
1 99.74 0.0000
39
20 140 100 100 100 33.34 28.57 66.67 100 100 99.83
TABLE 2: Extracted performance metric values from simulation with Non-Linear Kernel.
C CAR Time Gama
SVM
No
Back Land Pod Phf Rootkit Imap Dict NMap FPRR
1 99.01
0.0001
7
0.5 972 99.22 0 97.96 0 0 0 100 100 99.15
1 99.33
0.0000
79
2 511 99.45 0 97.96 0 0 0 100 100 99.77
1 99.31
0.0001
22
3 443 99.45 0 97.96 0 0 0 100 100 99.72
1 99.63
0.0000
54
10 326 99.84 100 97.96 0 0 66.67 100 100 99.89
2 99.72
0.0000
61
10 269 99.92 100 97.96
33.3
4
42.86 66.67 100 100 99.83
4 99.79
0.0000
43
10 246 100 100 97.96
33.3
4
42.86 66.67 100 100 99.94
8 99.79
0.0000
32
10 224 100 100 97.96
33.3
4
42.86 66.67 100 100 99.94
1
6
99.82
0.0000
32
10 211 100 100 97.96
66.6
7
42.86 66.67 100 100 99.94
1 99.60
0.0000
61
15 369 99.84 100 97.96 0 0 66.67 100 100 99.83
TABLE 3: Extracted performance metric values from simulation with Non-Linear Kernel.
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 4: Results of performance metrics for GA-Based Algorithms [12].

5. CONCLUSION AND FUTURE WORKS
In this paper a SVM based system is presented which can classify the IDS alerts with high
accuracy and reduce number of false positive alerts considerably. Also the system is able to
identify the attack types of the alerts more accurate in little time slice.
It seems to be useful using SVM to correlate alerts to discover attack sequences so this idea is a
future work of this paper.
6. REFERENCES
[1] Debar, H., M. Dacier, and A. Wespi, Towards a taxonomy of intrusion-detection systems.
Computer Networks, 1999. 31(8): p. 805-822.
[2] Cortes, C. and V. Vapnik, Support-vector networks. Machine learning, 1995. 20(3): p. 273-
297.
[3] Julisch, K., Clustering intrusion detection alarms to support root cause analysis. ACM
Transactions on Information and System Security (TISSEC), 2003. 6(4): p. 443-471.
[4] Maheyzah, S.Z., Intelligent alert clustering model for network intrusion analysis. Journal in
Advances Soft Computing and Its Applications (IJSCA), 2009. 1(1): p. 33-48.
[5] Wang, J., H. Wang, and G. Zhao. A GA-based Solution to an NP-hard Problem of
Clustering Security Events. 2006. IEEE.
[6] DARPA 2000 Intrusion Detection Evaluation Datasets, M.L. Lab., Editor. 2000.
[7] Cuppens, F. Managing alerts in a multi-intrusion detection environment. 2001.
[8] MIRADOR, E. Mirador: a cooperative approach of IDS. in European Symposium on
Research in Computer Security (ESORICS). 2000. Toulouse, France.
[9] Debar, H. and A. Wespi. Aggregation and Correlation of Intrusion-Detection Alerts. in
Proceedings of the 4th International Symposium on Recent Advances in Intrusion
Detection. 2001.
[10] Ahrabi, A.A.A., et al., A New System for Clustering and Classification of Intrusion Detection
System Alerts Using Self-Organizing Maps. International Journal of Computer Science and
Security (IJCSS), 2011. 4(6): p. 589.
[11] Kohonen, T., Self-Organized Maps. 1997, Science Berlin Heidelberg: Springer series in
information.
[12] Bahrbegi, H., et al. A new system to evaluate GA-based clustering algorithms in Intrusion
Detection alert management system. 2010. IEEE.
[13] Krovi, R. Genetic algorithms for clustering: a preliminary investigation. 1992. IEEE.
[14] Krishna, K. and M. Narasimha Murty, Genetic K-means algorithm. Systems, Man, and
Cybernetics, Part B: Cybernetics, IEEE Transactions on, 1999. 29(3): p. 433-439.
[15] Fuyan, L., C. Chouyong, and L. Shaoyi. An improved genetic approach. 2005. IEEE.
[16] Lu, Y., et al. FGKA: a fast genetic K-means clustering algorithm. 2004. ACM.
[17] Di Nuovo, A.G., V. Catania, and M. Palesi. The hybrid genetic fuzzy C-means: a reasoned
implementation. in International Conference on Fuzzy Systems. 2006. World Scientific and
Engineering Academy and Society (WSEAS).

[18] Ahrabi, A.A.A., et al., Using Learning Vector Quantization in IDS Alert Management
System. International Journal of Computer Science and Security (IJCSS), 2012. 6(2): p. 1-
7.
[19] Kohonen, T., Learning vector quantization, in M.A. Arbib (ed.), The Handbook of Brain
Theory and Beural Networks. 1995: MIT Press.
[20] Njogu, H.W., et al., A comprehensive vulnerability based alert management approach for
large networks. Future Generation Computer Systems, 2013. 29(1): p. 27-45.
[23] Brugger, S.T. and J. Chow, An Assessment of the DARPA IDS Evaluation Dataset Using
Snort, D. UC Davis Technical Report CSE-2007-1, CA, Editor. 2007.
[24] Vapnik, V.N., The nature of statistical learning theory. 2000: Springer-Verlag New York Inc.
[25] Webb, A.R., Statistical pattern recognition. Second Edition ed. 2002, Malvern UK: Wiley.
[26] Matlab, www.mathworks.com/products/matlab/, Editor. 2009, Mathworks.
[27] Ma, J., Y. Zhao, and S. Ahalt, OSU SVM classifier matlab toolbox (ver 3.00). Pulsed Neural
Networks, 2002.

Managing Intrusion Detection Alerts Using Support Vector Machines

Recommended

More Related Content

What's hot (16)

Similar to Managing Intrusion Detection Alerts Using Support Vector Machines (20)

Recently uploaded (20)

Managing Intrusion Detection Alerts Using Support Vector Machines