Machine Learning Based Botnet Detection

Machine Learning Based Botnet Detection

Vaibhav Nivargi Mayukh Bhaowal Teddy Lee
{vnivargi, mayukhb, tlee21}@cs.stanford.edu
CS 229 Final Project Report

I. INTRODUCTION these concerted efforts, Botnets remain an unsolved
problem for the online community.
A Botnet [1] is a large collection of compromised
machines, referred to as zombies [2], under a
common Command-and-Control infrastructure
(C&C), typically used for nefarious purposes.
Botnets are used in a variety of online crimes
including, and not limited to, large scale DDoS
attacks, Spam, Click Fraud, Extortion, and Identity
theft. The scale and geographical diversity of the
machines enlisted in a Botnet, coupled with easily
available source code, and support from
communities, as well as mercenary Botmasters
providing Botnet services for rent, have resulted in
Botnets becoming a highly sophisticated and
effective tool for committing online crime in recent
times [3][4]. Botnets with thousands and millions
of nodes have been observed in the wild, with
newer ones being observe every day [10].
The lifecycle of a Botnet is depicted in
Fig.1. The initial enlisting occurs by exploiting a
known vulnerability in the Operating systems
running on the machines using a wide variety of Fig 1. Botnet in action
mechanisms like worms, Trojans, P2P file sharing
networks, and exploits of common Windows II. DATA
vulnerabilities, etc. Once compromised, the bot is
programmed to connect to a central location We had two separate data sets to collect for the
(typically an IRC [11] server), where the Botmaster purpose of our experiments. The first set included a
could login and issue commands to the logged in large number of binaries/executables labeled as
bots. This mechanism essentially means that the botnet binaries and benign binaries. We acquired
communication is free, as broadcast is taken care of the botnet binaries from the computer science
by the IRC channel. Most Bots additionally ship department at John Hopkins University. This is the
with multiple IRC and DNS addresses, meaning same data they used for botnet related work [18].
taking down one such IRC channel does not in As far as the benign binaries are concerned we
general impair the activities of the Botnet. randomly picked them up from Unix and windows
Originally, most techniques to thwart machines, namely from /usr/bin and from windows
Botnets have been reactive, reducing their system32 directory. This data was comprehensive
effectiveness significantly, e.g. using Honeypots to and well represented.
trap and study malware, etc. Of late, significant We also needed labeled IRC logs for our
research has been made into the dynamics of a experiments. While benign IRC logs are easily
Botnet, and more proactive techniques have been available on the web, botnet affected IRC logs are
suggested [5] [6] [7] [8]. A wide variety of features not readily available because of privacy and legal
and watermarks for network activity are employed issues. The benign IRC logs were collected from
to predict Botnet activity, including TCP syn several publicly available IRC logs from IRC
scanning, DNS monitoring, and extensive models channels like Wikipedia, linode, and
of Botnet attack and propagation [9]. Despite all kernelnewbies. These represent a diverse collection
of logs, with different purposes. Some of these logs

also have automated bots for channel maintenance We used supervised learning techniques on groups
operations. of benign executables vs. Botnet executables.
Obtaining malicious IRC logs proved to be
extremely hard. There are several independent 3.1.1 Features
Security forums who actively track botnets [12].
They monitor and mine Botnet IRC logs for We are focusing on binary profiling, and hex
analysis and mitigation. Due to privacy and security dumps for feature extraction. Identifying the strain
issues, we were unable to obtain this data, which of these binaries might also give an insight about
clearly represents a rich set of real world Botnet the location of the Command-and-Control center
IRC logs, and which would have definitely and about the botnet capabilities in general. We
provided more qualitative as well as quantitative used n-grams (more specifically 4-grams) of
results. Several other potential sources setup their hexdump of the binaries as our features. For
own private infrastructure for collecting such example if the hexdump is ff 00 12 1a 32, our
training data [5]. features will be ff 00 12 1a and 0012 1a 32. We
Nevertheless we acquired data from extracted around more than a million features. We
Northwestern University where the department of then used chi-square to select around 10,000 most
CS is conducting research on wireless overlay informative features. For each feature, we find its
network architectures and botnet detection [7]. The value as follows:
data regarding botnet IRC logs was not
comprehensive in the sense that it was IRC traffic feature=00121a32 feature≠00121a32
over a small amount of time. A larger and more class=botnet A C
comprehensive dataset could have established our class≠botnet B D
results and hypothesis more conclusively.
N ( AD − CB) 2
III. APPROACHES χ 2 (botnet, f ) =
( A + C )(B + D)( A + B)(C + D)
.
We tried a 2 stage approach to solve this issue. Now we select the top 10,000 features with the
These methods are complementary and we can highest chi-square scores.
combine them for better results. They are as
follows: 3.1.2 Classification
3.1 Botnet Binary Detection We used several classification algorithms to
classify the binaries into malicious/benign. The
There are several stages in the lifecycle of a Botnet models we used are as follows:
where a Machine learning based solution can be 1. Multinomial Naïve Bayes
deployed to thwart its effectiveness. During an 2. linear SVM
early stage, a Binary detection and classification 3. kNN
system can warn of a potentially malicious This makes use of similarity metrics (e.g.
executable which might endanger a host machine. cosine similarity) to find k nearest neighbors
There has been some work already in this area [7] and classifies the point under consideration
and we leverage on top of their work to classify into the majority class of its k nearest neighbor
Botnet executables which propagate as worms on [15]. We used k=5.
networks scanning for vulnerable machines for 4. Logistic Regression
infecting them, and enlisting them into the Bot 5. Multiboost Adaboost - Class for boosting a
network. classifier using the MultiBoosting
Unlike Virus scanners like Norton AV, or method. MultiBoosting is an extension to the
McAfee, a Machine learning solution can perform highly successful AdaBoost technique for
this classification without the need for explicit forming decision committees. MultiBoosting
signatures. Identifying such binaries without can be viewed as combining AdaBoost with
explicit signatures needs recognizing common wagging. It is able to harness both AdaBoost's
features and correlations between these binaries, high bias and variance reduction with
e.g. a Botnet executable will be a self-propagating wagging's superior variance reduction. Using
worm with a simple IRC client built in. Presence or C4.5 as the base learning algorithm, Multi-
absence of such a feature is an indicator that such a boosting is demonstrated to produce decision
binary might potentially be a Botnet executable. committees with lower error than either

AdaBoost or wagging significantly more often attributes not yet considered in the path from
than the reverse over a large representative the root.
cross-section of UCI data sets. It offers the
further advantage over AdaBoost of suiting For the purpose of our experiments we did not
parallel execution. See [17] for more details. make use of any separate testing data set. To stay
unbiased we used 10 fold cross validation to get the
6. J48 Decision tree -This is an entropy based results. We used off the shelf softwares such as
approach for generating a pruned or unpruned weak[19] and libsvm[13] for experimental results.
C4.5 decision tree. For more information see
[16]. In general, if we are given a probability 3.2 IRC log based detection
distribution P = (p1, p2, .., pn) then the
Information conveyed by this distribution, also IRC has played a central role in the
called the Entropy of P, is: simplicity and effectiveness of a Botnet. Using a
public communication channel, the Botmaster can
I ( P) = −( p1 * log( p1 ) + p2 * log( p2 ) + ... + pn * log( pn )) use a simple command interface to communicate
with a huge number of compromised zombie nodes,
instructing them to carry out his orders.
If a set T of records is partitioned into disjoint There are two phases to this approach:
exhaustive classes C1, C2, .., Ck on the basis First, to separate IRC traffic from other traffic. This
of the value of the categorical attribute, then is a reasonably solved problem [5][6]. The second
the information needed to identify the class of step comprises of identifying Botnet traffic in the
an element of T is Info(T) = I(P), where P is IRC traffic. Hence the problem now boils down to
the probability distribution of the partition (C1, a text classification problem.
C2, .., Ck): To be able to differentiate a benign IRC
log from an IRC log manifested with Botnet
activity, we used features involving both dialogues
and IRC commands. Then using these features, we
| C1 | | C 2 | | Ck | experimented with a variety of machine learning
P=( , ,..., ) algorithms on them. In particular, we ran
|T | |T | |T |
algorithms such as Naïve Bayes, SVM, J48
decision trees, kNN, etc. with 10 fold cross
If we first partition T on the basis of the value validation. The main categories of features we used
of a non-categorical attribute X into sets T1, T2, included:
.., Tn then the information needed to identify
the class of an element of T becomes the Number of users: An IRC channel with Botnet
weighted average of the information needed to activity should contain an unusually large
identify the class of an element of Ti, i.e. the number of users.
weighted average of Info(Ti): Mean / variance of words per line and
characters per word in dialogues: Bots usually
n
| Ti | do not produce dialogues that resemble human
Info( X , T ) =∑ i =1 | T |
* Info(Ti ) dialogue.
Number and frequency of IRC commands: We
have noticed through examination of the logs
Consider the quantity Gain(X,T) defined as that there tends to be a large number of IRC
commands at small intervals in Botnet
Gain( X , T ) = Info(T ) − Info( X , T ) manifested logs. One possible explanation
would be the immense number of joins and
This represents the difference between the exits from the result of accommodating a huge
information needed to identify an element of T number of users in one channel.
and the information needed to identify an Number of lines, dialogues, and commands: In
element of T after the value of attribute X has a benign IRC log, the number of dialogues
been obtained, that is, this is the gain in should be much greater than the number of
information due to attribute X. We can use this commands. And as mentioned above, a Botnet
notion of gain to rank attributes and to build manifested log tends to contain an immense
decision trees where at each node is located the number of IRC commands.
attribute with greatest gain among the

IV. RESULTS AND EVALUATION Model Accuracy F1 Kappa
score score
4.1 Botnet Binary Detection
NB .70 .457 .705
The results obtained from the botnet binary based
detection approach are summarized in Fig. 2. Linear SVM .976 .945 .97
Clearly all the models performed reasonably well.
Special mention must be made about Naïve Bayes
which performed remarkably well although it is J48 – decision .992 .982 .99
one of the simplest of models. SVM performed tree
good too. However some models like kNN gave
an accuracy of 96.4 which was lower than that of kNN .992 .982 .991
others. We will discuss about these results in
details in the discussion section. In particular our
evaluation metric included accuracy, F1 score and Logistic .989 .974 .987
kappa measure.
MultiboostAB .967 .926 .963

# Correctly predicted data po int s
Accuracy = Fig. 3 Performance of IRC based detection
# Total data po int s

4.2 IRC log based detection
2 * Pr ecision * Re call
F1 =
Pr ecision + Re call The results obtained from the IRC log
based detection approach are summarized in Fig. 3.
A large chunk of the problem here involves text
Observed Agreement − Chance Agreement classification, and all algorithms, barring Naïve
Kappa = Bayes, perform encouragingly well.
Total Observed − Chance Agreement
V. DISCUSSION

In the binary based classification approach, most of
the linear classifiers such as NB and linear SVM
performed remarkably well. kNN however
Model Accuracy F1 Kappa performed relatively bad. This can be explained on
score score the basis of linearity of the data set. kNN is a non-
.982 .981 .963 linear model and hence it has less bias and a high
NB
variance unlike NB which is linear and has high
bias and less variance. Given our data set was
Linear SVM .982 .981 .963 linear, it was therefore no surprise that kNN
exhibited a higher generalization error compared to
J48 – decision .976 .975 .95 NB or linear SVM.
tree In the IRC log based approach, most
classifiers performed similarly, except for Naïve
Bayes, which was significantly worse. One
kNN .964 .963 .926 possible explanation would be that there are
dependencies in the features. These dependencies
Logistic .982 .9815 .963 may be resolved with more training data. Since the
set of IRC logs we obtained was not very large and
comprehensive, it is possible that currently
MultiboostAB .976 .975 .951 correlated features are not actually correlated in a
larger dataset.
Fig. 2 Performance of binary based detection The other classifiers performed very well,
all having accuracies greater than .96. Since SVM

and logistic regression had the best accuracies, F1 JHU for botnet binary datasets. We further
scores, and Kappa scores, it is likely that our acknowledge our gratitude to Prof. Andrew Ng and
dataset is linear. the TAs of cs229 for their help and support.
In the logs that we collected with Botnet
activity, there weren’t strong attempts of log VIII. REFERENCES
obfuscation by Botmasters. But suppose that there
were strong attempts to produce human-like [1] http://en.wikipedia.org/wiki/Botnet
dialogues by the bots. Then to potentially fool our [2] http://en.wikipedia.org/wiki/Zombie_computer
classifiers, the dialogues must produce similar [3]http://securitywatch.eweek.com/exploits_and_at
averages and variances in terms of words per line tacks/everydns_opendns_under_botnet_ddos_attac
and characters per word. k.html
Pushing the above scenario to the extreme, [4]http://it.slashdot.org/article.pl?sid=06/10/17/002
suppose that the bots produce perfect dialogues. In 251
this case, our classifier should still work due to one [5] W. Timothy Strayer, Robert Walsh, Carl
fundamental difference between benign logs and Livadas, and David Lapsley. Detecting Botnets
logs manifested with Botnet activity – there is an with Tight Command and Control. To Appear in
immense number of join and exit commands in a 31st IEEE Conference on Local Computer
Botnet infested channel. Networks (LCN'06). November 2006
[6] Using Machine Learning Techniques to Identify
Botnet Traffic, Carl Livadas, Robert Walsh, David
VI. CONCLUSION Lapsley, W. Timothy Strayer
[7] Yan Chen, “Towards wireless overlay network
In this course project we aimed to architectures”
highlight a major problem plaguing the Internet [8] Binkley- An Algorithm for Anomaly-based
today and a potential solution to the problem by Botnet Detection .
employing novel Machine Learning techniques. [9] E. Cooke, F. Jahanian, and D. McPherson. The
Our results are encouraging, indicating a definite zombie roundup: Understanding, detecting, and
advantage of using such techniques in the wild for disrupting botnets. In USENIX SRUTI Workshop,
alleviating the problem of Botnets. pages 39–44, 2005.
Conversion of the problem in the Text [10]http://www.vnunet.com/vnunet/news/2144375/
classification domain has opened up new avenues botnet-operation-ruled-million
in feature extraction for our solution. Possible [11] http://www.irc.org/tech_docs/rfc1459.html
future enhancements include adding additional [12] www.shadowserver.org
features to the IRC traffic classification engine. [13] http://www.csie.ntu.edu.tw/~cjlin/libsvm/
Also to solve the problem of training data which, [14] Detecting Botnets with Tight Command and
in retrospect, seems to be the critical to the success Control, Carl Livadas, Robert Walsh, David
of such a project, we propose setting up a farm of Lapsley, W. Timothy Strayer.
Virtual machines as a testbed for collecting and [15] Aha, D., and D. Kibler (1991) "Instance-
mining our own logs for training purpose. based learning algorithms", Machine Learning,
Lastly, the end goal of such a tool should vol.6, pp. 37-66.
be to provide proactive help to system [16] Ross Quinlan (1993). "C4.5: Programs for
administrators. In this regard, it is easy to envisage Machine Learning", Morgan Kaufmann Publishers,
an application which can run on a Gateway or a San Mateo, CA.
Router, and look at Network flow traffic, and [17] Geoffrey I. Webb (2000). "MultiBoosting: A
classify it on-the-fly enabling automated Technique for Combining Boosting and Wagging".
blacklisting of involved machines in the network. Machine Learning, 40(2): 159-196, Kluwer
Academic Publishers, Boston
[18] Moheeb Abu Rajab et.al (2006)”A
Multifaceted approach to understanding the botnet
VII. ACKNOWLEDGEMENT phenonmenon”
[19] http://www.cs.waikato.ac.nz/ml/weka/
We take this opportunity to thank Elizabeth Stinson
of Stanford Security Group for her help and
support. We also extend our acknowledgements to
Yan Chen of northwestern university for providing
us with botnet IRC logs and to Andreas Terzis of

Machine Learning Based Botnet Detection

More Related Content

What's hot (20)

Similar to Machine Learning Based Botnet Detection (20)

More from butest (20)

Machine Learning Based Botnet Detection