Protocol Type Based Intrusion Detection Using RBF Neural Network

Aslıhan Özkaya & Bekir Karlık
International Journal of Artificial Intelligence and Expert Systems (IJAE), Volume (3) : Issue (4) : 2012 90
Protocol Type Based Intrusion Detection
Using RBF Neural Network
Aslıhan Özkaya aozkaya@mevlana.edu.tr
Faculty of Engineering/Computer Engineering Department
Mevlana University
Konya, 42003, TURKEY
Bekir Karlık bkarlik@mevlana.edu.tr
Faculty of Engineering/Computer Engineering Department
Mevlana University
Konya, 42003, TURKEY
Abstract
Intrusion detection systems (IDSs) are very important tools for information and computer se-
curity. In IDSs, the publicly available KDD’99, has been the most widely deployed data set
used by researchers since 1999. Using a common data set has provided to compare the re-
sults of different researches. The aim of this study is to find optimal methods of preprocessing
the KDD’99 data set and employ the RBF learning algorithm to apply an Intrusion Detection
System.
Keywords: RBF Network, Intrusion Detection, Network Security, KDD Dataset.
1. INTRODUCTION
With the growth in the use of computer and internet, the number of computer and network
attacks has increased. Therefore many companies and individuals are looking for solutions
and deploying software’s and systems such as intrusion detection systems (IDSs) to over-
come with the network attacks. Due to the high need of such systems, many researchers’
attentions are attracted by IDS [1-4].
KDDCUP'99 is the mostly widely used data set for the evaluation of intrusion detection sys-
tems [5-8]. Tavallaee et al. [5] examined and questioned the KDDcup’99 data set, and revised
it by deleting the redundant records and applied 7 learners on the new data set. The seven
learners are J48 decision tree learning, Naive Bayes, NBTree, Random Forest, Random
Tree, Multilayer Perceptron (MLP), and Support Vector Machine (SVM). They also labeled
each record with its difficulty and present it publicly on their website. Sabhnani and Serpen [6]
evaluated the performance of pattern recognition and machine learning algorithms on KDD’99
data set. In their paper the following algorithms are tested; MLP, Gaussian classifier, K-
means, nearest cluster algorithm, incremental RBF, Leader algorithm, Hyper sphere algo-
rithm, Fuzzy ARTMAP and C4.5 decision tree. They mainly focused on comparing the per-
formances of the applied classifiers for the attack categories. Bi et al. [7] picked 1000 records
from KDDcup’99. They used Radial Basis Function (RBF) Network on the selected data after
preprocessing it. Sagiroglu et al. [8] applied Leverberg Marquardt, Gradient Descent, and
Resilient Back-propagation on the KDD’99 data set.
The other machine learning algorithms are also used for intrusion detection. Yu and Hao [9]
presented an ensemble approach to intrusion detection based on improved multi-objective
genetic algorithm. O. A. Adebayo et al. [10] have presented a method that uses Fuzzy-
Bayesian to detect real-time network anomaly attack for discovering malicious activity against
computer network. Shanmugavadivu and Nagarajan [11] presented fuzzy decision-making
module to build the system more accurate for attack detection using the fuzzy inference ap-
proach. Ahmed and Masood [12] proposed a host based intrusion detection architecture using
RBF neural network which obtained better detection rate and very low training time as com-
pared to other machine learning algorithms.

In this study, the KDD’99 data set has been pre-processed and divided into three sections
according their protocol type; TCP, UDP and ICMP. Conversion of string to numerical value is
applied in three different ways and is saved as three different data sets. RBF neural network
learning algorithm is used for each data set.
2. DATA SET DESCRIPTION AND PREPROCESSING
2.1. KDD’99 Data Set
In our experiments we have used the KDD’99 data set which has been developed based on
the data captured in DARPA’98 [13]. The KDD’99 data set (corrected version) has over 1
million training data and over 300 thousands of test data. Each data consists of 41 attributes
and one target (see Figure 1). Targets indicate the attack names. The data set covers over 30
different attack types as outputs which belong to one of four major categories; Denial of Ser-
vice, User to Root, Remote to Local, and Probing Attacks (see Table 1) [4].
FIGURE 1: Sample data of KDDcup
2.2. Deleting Repeated Data
We used MATLAB on a PC with 4 GB of memory and 2.27 GHz of processing speed. Be-
cause of the limited memory and speed of the PC we decided to decrease the number of data
of the training sets to around 6,000. Therefore repeated data has been deleted. After this
process 614,450 of training and 77,290 of testing data was left.
2.3. Dividing Data into Three Sections
As shown in Figure 1, one of the attributes is the protocol type which is TCP, UDP or ICMP.
We divided both training and testing data into these three protocol types in order to train and
test our data separately.
Table 2 shows the number of remaining data after repeated data has been deleted. The num-
ber of training data for TCP and UDP is still large. Therefore some number of data was de-
leted randomly. The data to be deleted were chosen mostly from “normal” labeled data.
There were also some attacks in testing data set that were not in the training data set. Since
RBF is a supervised learning technique, we had to train the network for all attacks which are
going to be tested. Therefore we copied some of these attacks into the training data set. But
the testing data sets were untouched (see Table 3).

Category Attack Name TEST TRAIN Category Attack Name TEST TRAIN
apache2. 0 794
back. 2,002 1,098
land. 17 9
ftp_write. 8 3 mailbomb. 0 5
guess_passwd. 53 4,367 neptune. 204,815 58,001
httptunnel. 0 158 pod. 40 87
imap. 12 1 processtable. 0 759
multihop. 6 18 smurf. 227,524 164,091
named. 0 17 teardrop. 199 12
phf. 3 2 udpstorm. 0 2
sendmail. 0 17 buffer_overflow. 5 22
snmpgetattack. 0 7,741 loadmodule. 2 2
snmpguess. 0 2,406 perl. 2 2
warezmaster. 20 1,602 ps. 0 16
worm. 0 2 rootkit. 0 13
xlock. 0 9 sqlattack. 0 2
xsnoop. 0 4 xterm. 0 13
ipsweep. 7,579 306 TOTAL 1,048,575 311,029
mscan. 0 1,053
nmap. 2,316 84
portsweep. 2,782 354
saint. 0 736
satan. 5,393 1,633
ProbingRemotetoLocal(R2L)
UsertoRoot(U2R)
Normal
normal. 595,797 60,593
DenialofService(DoS)
TABLE 1: Attack categories for test and training data sets
Protocol Name:
TRAIN TEST TRAIN TEST TRAIN TEST
Normal 529,517 43,908 28,435 3,770 1,324 233
Attack 50,499 27,214 866 922 3,809 1,242
Total 580,016 71,122 29,301 4,692 5,133 1,475
ICMPUDPTCP
TABLE 2: Data information after separating it into three different protocol types.
Protocol Name:
TRAIN TEST TRAIN TEST TRAIN TEST
Normal 2,698 43,908 5,134 3,770 1,325 233
Attack 3,302 27,214 942 922 3,838 1,242
Total 6,000 71,122 6,076 4,692 5,163 1,475
UDP ICMPTCP
TABLE 3: Data information after deleting some data randomly and copy some attacks from test to
train data set
2.4. Normalization
In order to normalize the data, we need to make sure that all values are in numerical formats.
There are three inputs (attributes) and one output given in string formats. One of these
attributes is the protocol name. Since the data is divided according their protocol names,
there is no need to convert the protocol types to numeric values. We deleted the column
which belongs to the protocol name, since one set of data has always the same protocol
name. The output has been converted to 1 if it is an attack and to 0 (zero) if it is a normal
communication data. The other two attributes are the service and flag name. They are con-
verted to numerical values with respect to their frequency in the test set. We applied three
different conversion techniques. We differentiated these techniques by naming them as Type-
A, Type-B and Type-C.

Flag Frequency Type-A Type-B Type-C
SF 6765 1 11 10
S0 3986 2 10 6
REJ 1488 3 9 2
RSTR 633 4 8 5
RSTO 307 5 7 3
S3 272 6 6 9
SH 182 7 5 11
S1 58 8 4 7
S2 29 9 3 8
RSTOS0 25 10 2 4
OTH 4 11 1 1
TABLE 4: Conversions of Flag Names to Numerical Values (for TCP Data set)
Service
Name
Frequency Type-A Type-B Type-C
Service
Name
Frequency Type-A Type-B Type-C
private 3156 1 57 40 rje 26 30 33 42
http 3012 2 56 17 daytime 25 31 24 6
telnet 1669 3 55 50 netbios_dgm 25 32 25 29
ftp 910 4 54 13 supdup 25 33 26 48
other 864 5 53 35 uucp_path 25 34 27 53
ftp_data 821 6 52 14 bgp 24 35 20 2
smtp 765 7 51 44 ctf 24 36 21 5
finger 507 8 50 12 netbios_ssn 24 37 22 31
pop_3 401 9 49 38 whois 24 38 23 55
imap4 227 10 48 19 csnet_ns 23 39 17 4
auth 177 11 47 1 name 23 40 18 28
sunrpc 113 12 46 47 vmnet 23 41 19 54
IRC 110 13 45 20 hostnames 22 42 15 16
time 88 14 44 51 Z39_50 22 43 16 57
domain 52 15 43 8 nntp 18 44 13 34
remote_job 40 16 42 41 pm_dump 18 45 14 36
sql_net 39 17 40 45 ldap 15 46 12 24
ssh 39 18 41 46 uucp 10 47 11 52
X11 32 19 39 56 login 9 48 10 26
discard 29 20 36 7 nnsp 7 49 7 33
echo 29 21 37 9 printer 7 50 8 39
systat 29 22 38 49 shell 7 51 9 43
gopher 28 23 34 15 kshell 6 52 6 23
link 28 24 35 25 courier 5 53 3 3
iso_tsap 26 25 28 21 exec 5 54 4 11
mtp 26 26 29 27 http_443 5 55 5 18
netbios_ns 26 27 30 30 efs 4 56 2 10
netstat 26 28 31 32 klogin 3 57 1 22
pop_2 26 29 32 37
TABLE 5: Conversions of Service Names to Numerical Values (for TCP Data)
Service
Name
Frequency Type A Typle B Type C
domain_u 3679 1 5 1
ntp_u 1373 2 4 3
private 795 3 3 5
other 152 4 2 2
tftp_u 0 5 1 4
TABLE 6: Conversions of Service Names to Numerical Values (for UDP Data)

In Type-A, we gave the highest number to the attribute with most frequency and 1 with less
frequency. We did this in the opposite way for Type-B, and random numerical values were
given in Type-C.
Service
Name
Frequency Type A Typle B Type C
eco_i 2990 5 1 1
ecr_i 1727 4 2 5
urp_i 270 3 3 2
urh_i 146 2 4 4
tim_i 0 1 5 3
TABLE 7: Conversions of Service Names to Numerical Values (for ICMP Data)
There is only on Flag name in ICMP and UDP data sets; therefore the columns belong to the
flag names are deleted for both ICMP and UDP. There were also some other columns with
only one value. These columns (inputs) are also deleted because they have no influence on
the outputs. The final number of inputs and outputs of the data sets can be seen in Table 8.
Protocol Name: TCP UDP ICMP
Input # 31 20 18
Output # 1 1 1
TABLE 8: Number of Output and Input after preprocessing the data sets
After converting text to integer and deleting columns with same data, the data sets are norma-
lized.
3. RBF NETWORK
Radial Basis Function (RBF) Network is a type of Artificial Neural Network for supervised
learning [14]. It uses RBF as a function which is usually Gaussian and the outputs are in-
versely proportional to the distance from the center of the neuron [15]. The traditional RBF
function network can be seen in Figure 2. MATLAB provides functions to implement RBF
Network within their Neural Network Toolbox. The training function newrb() and simulation
function sim() is used to train and test the network [15-16].
FIGURE 2: A single layer radial basis function network

4. EXPERIMENTS
The experiments are applied for all three types of string to integer conversations to see if
there is any difference. For all trainings the maximum number of neurons is set as 1000.
4.1. Training Results
Training results can be seen in Table 9, 10 and 11. The results are shown as mean squared
error (MSE) which represents the performance (or accuracy).
The best training results are Type-C for TCP, Type-A for UDP and Type-B for ICMP.
# of Neurons Type-A Type-B Type-C
50 0.02702 0.02718 0.02985
100 0.01540 0.01575 0.01648
150 0.01127 0.01097 0.01275
200 0.00900 0.00869 0.00927
250 0.00772 0.00722 0.00680
500 0.00321 0.00335 0.00295
750 0.00165 0.00157 0.00151
1000 0.00101 0.00097 0.00089
TCP TRAINING
TABLE 9: Training results (MSE) of the TCP data
set
TABLE 10: Training results (MSE) of the UDP
data set
TABLE 11: Training results (MSE) of the ICMP data set
The training performances are plotted to set the results of one type of conversion against the
other types of conversion (see Figure 3, 4, and 5). It can be seen that the learning perfor-
mances for each type is very close to each other.
FIGURE 3: Graphical training results of the TCP data set

FIGURE 4: Graphical training results of the UDP data set
FIGURE 5:Graphical training results of the ICMP data set
4.2. Testing Results
The best performance is obtained with Type-C conversion of all three data sets. The MSE
and FAR values are 95.65%, 79.39%, 62.96% and 2.6%, 4.72%, 7.85% for TCP, UDP and
ICMP respectively.
Figure 6 and Figure 7 show the comparison of the performances and False Alarm Rates for
TCP, UDP and ICMP testing data sets with their three different Type of conversions (Type-A,
Type-B and Type-C).
Type-A Type-B Type-C
Performance 90.86% 94.28% 95.65%
False Alarm 3.45% 3.38% 2.60%
Performance 61.42% 65.09% 63.96%
False Alarm 8.78% 10.29% 7.85%
Performance 88.95% 83.46% 79.39%
False Alarm 16.31% 15.88% 4.72%
UDPICMPTCP
TABLE 12: Testing Results for TCP, UDP and ICMP data sets.

FIGURE 6: Testing result of Performances.
FIGURE 7: Testing results of False Alarm Rates (FARs)
False alarm rates of all type of conversions have been observed similar for both TCP and
UDP testing datasets. The FAR results for ICMP testing dataset have an appreciable amount
of differences. It is observed that FARs are over 15% for type-A and type-B while it is less
than 5% for type-C.
According to experimental results, false alarms are always the highest percentage for type-A
and type-B data sets. This shows that converting strings to numbers with respect to their fre-
quency may not be a good solution.
Learning and testing the TCP dataset gives good results and can still be improved, while the
results for UDP and ICMP datasets are very poor. More training data or more attributes may
improve the results.
In this paper the overall MSE and FAR values are calculated as 93.42% and 2.95% respec-
tively. These results are better than the results in some other papers where different methods
have been applied. For instance in [5] the performance values are 81.66%, 92.79%, 92.59%,
92.26% and 65.01% with Naïve Bayes, Random Forest, Random Tree, Multi-Layer Percep-
tron, and SVM respectively. Again in the same paper the performance values of some other
methods (J48 and NB Tree) are very close to our overall results which are 93.82% and
93.51% respectively. In [7] the performance is 89% and FAR is 11% with RBF neural network.

5. CONCLUSION AND DISCUSSION
In this study, the most widely used data set (KDD’99) is pre-processed. Some duplicated data
is deleted then training and testing data is divided into three sections according the protocol
types. Afterwards strings in the data sets are converted to numerical values using three dif-
ferent techniques as Type-A, Type-B and Type-C. All preprocessed data sets are trained and
tested with RBF network using MATLAB toolbox. It is experimented that the preprocessing
phase plays an important role on the performance of the learning system.
It is also observed that applying learning algorithms on divided data (with respect to their pro-
tocol types) enables better performance.
As mentioned in the testing results section, the accuracy of testing results is more satisfied
than the literature studies. However this proposed learning algorithm and alternative string to
integer converting techniques need more research to find optimal solutions.
6. REFERENCES
[1] K. Ilgun, R.A Kemonerer and P.A Porras, “State Transition Analysis: A Rule Based
Intrusion Detection Approach”, IEEE Transaction on Software Engineering, Vol.21(3),
March 1995, pp.181-199.
[2] S. Capkun. Levente Buttyan. “Self-Organized Public Key Management For Mobile Ad
Hoc Networks”, IEEE Transactions on Mobile Computing, Vol. 2(1), January -March
2003, pp. 52-64.
[3] Yao, J. T., S.L. Zhao, and L.V. Saxton, “A Study On Fuzzy Intrusion Detection”, In Pro-
ceedings of the Data Mining, Intrusion Detection, Information Assurance, and Data
Networks Security, SPIE, Vol. 5812,pp. 23-30, Orlando, Florida, USA, 2005.
[4] 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, and M. A. Zissman, “Evaluating In-
trusion Detection Systems: The 1998 DARPA Off-Line Intrusion Detection Evaluation,”
in Proc. DARPA Inf. Survivability Confer. Exposition (DISCEX), Vol. 2, 2000, pp. 12–26.
[5] M. Tavallaee, E. Bagheri, W. Lu, and A. Ghorbani, “A Detailed Analysis of the KDD
CUP 99 Data Set,” in Proc. 2009 IEEE International Conference on Computational
Intelligence for Security and Defense Applications. pp. 53-58.
[6] M. Sabhnani and G. Serpen, “Application of Machine Learning Algorithms to KDD 1999
Cup Intrusion Detection Dataset within Misuse Detection Context”, International Confe-
rence on Machine Learning, Models, Technologies and Applications Proceedings, Las
Vegas, Nevada, June 2003, pp. 209-215.
[7] J. Bi, K. Zhang, X. Cheng, “Intrusion Detection Based on RBF Neural Network”, Infor-
mation Engineering and Electronic Commerce, 2009, pp. 357 - 360
[8] Ş.Sagiroglu, E. N. Yolacan, U. Yavanoglu, “Designing and Developing an Intelligent
Intrusion Detection System”, Journal of the Faculty of Engineering and Architecture of
Gazi University, Vol. 26 (2), June 2011, pp. 325-340.
[9] Yu Y, and Huang Hao, “An Ensemble Approach to Intrusion Detection Based on Im-
proved Multi-Objective Genetic Algorithm”, Journal of Software, Vol.18 (6), June 2007,
pp.1369-1378.
[10] O. Adetunmbi Adebayo, Zhiwei Shi, Zhongzhi Shi, Olumide S. Adewale, “Network
Anomalous Intrusion Detection using Fuzzy-Bayes", IFIP International Federation for
Information Processing, Vol. 228, 2007, pp. 525-530.
[11] R. Shanmugavadivu and N. Nagarajan, “An Anomaly-Based Network Intrusion Detec-
tion System Using Fuzzy Logic”, International Journal of Computer Science and Infor-
mation Security, Vol. 8 (8), November 2010, pp. 185-193.

[12] U. Ahmed and A. Masood, “Host Based Intrusion Detection Using RBF Neural Net-
works”, Emerging Technologies, ICET 2009, 19-20 Oct. 2009, pp. 48-51.
[13] The UCI KDD Archive, University of California, KDD Cup 1999 Data,
http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html, October 28, 1999, [Feb
2012].
[14] J. Mark and L. Orr, “Introduction to Radial Basis Function Networks”, Technical Report,
April 1996.
[15] Z. Caiqing, Q. Ruonan, and Q. Zhiwen, “Comparing BP and RBF Neural Network for
Forecasting the Resident Consumer Level by MATLAB,” International Conference on
Computer and Electrical Engineering, 2008 (ICCEE 2008), 20-22 Dec. 2008, pp.169-
172.
[16] A. Iseri and B. Karlık, “An Artificial Neural Networks Approach on Automobile Pricing”,
Expert Systems with Applications, Vol. 36 (2), March 2010, pp. 2155-2160.

Protocol Type Based Intrusion Detection Using RBF Neural Network

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Protocol Type Based Intrusion Detection Using RBF Neural Network (20)

More from Waqas Tariq (20)

Recently uploaded (20)

Protocol Type Based Intrusion Detection Using RBF Neural Network