A Dynamic Botnet Detection Model based on Behavior Analysis

Int. J. on Recent Trends in Engineering and Technology, Vol. 10, No. 1, Jan 2014

A Dynamic Botnet Detection Model based on
Behavior Analysis
K.Muthumanickam1, E.Ilavarasan2 , Sanjeev Kumar Dwivedi3
1,2,3

Research Scholar1, Associate Professor2, M.Tech. Student3
Department of Computer Science and Engineering, Pondicherry Engineering College, Pondicherry, India
Email: 1 kmuthoo@pec.edu , 2eilavarasan@pec.edu, 3sanjeevdwivedi131988@gmail.com

Abstract— Today different types of malware exist in the Internet. Among them one of the
malware is known as botnet which is frequently used for many cyber attacks and crimes in
the Internet. Currently botnets are the main rootcause for several illegal activities like
spamming, DDoS, click fraud etc. Botnets operate under the command and control(C&C)
infrastructure which makes its functioning unique. As long as the Internet exists botnet also
will exist. It can be used to perpetrate many Internet crimes. So fighting against them is a
challenging problem. The P2P-decentralized based botnets are more dangerous than
centralized botnets. In this paper a novel approach for the detection of P2P based botnet is
presented. The proposed approach for the detection of botnet in the network stream
analysis has been done in three phases. The first phase begins with the identification of P2P
node and the second phase deals with the clustering of the suspicious P2P node. Finally
botnet detection procedure has been applied which is based on stability of bots.
Experimental results show that the proposed approach detects more number of bots with
high accuracy.
Index Terms— Bots, P2P botnet, In-out degree, Clustering, flow records, Stability.

I. INTRODUCTION
Malicious malware can exploit vulnerabilities in the Internet computing Environment without user’s
knowledge. Basically malware includes viruses, worms, Trojan horses, and spyware that can be used to
gather information about a computer user and access to a system without their permission. It can appear in
the form of scripts, active content, code, or other software. Malware programs are divided into two classes:
first class of malwares needs a host program (viruses, Trojan horses, logic bombs, trapdoors) and second
class of malwares are independent programs (worms, zombie). Malwares are classified based on their
characteristics; some malwares do not replicate (activated by trigger) and others that produce copies of
themselves.
Botnet has become the most serious security threat on the current internet infrastructure. A botnet
(BotNetwork) is an interconnected collection of compromised infected computers (bots) which is remotely
controlled by its originator (called botmaster or botherder) under a common and control infrastructure [1].
Bot is a new type of malware which is designed for malicious activity. After the bot code has been installed
into a computer, the computer becomes a member of the botnetwork. Here all the bots are under the control
of BotMaster. So if bot exist in computer, it is not harmful until it receives command from BotMaster. After
receiving the command from BotMaster, it becomes dangerous for the system. These bots are not selfDOI: 01.IJRTET.10.1.18
© Association of Computer Electronics and Electrical Engineers, 2014

propagate from one system/network to other system/network [2]. A botnet enrolls its soldiers using social
engineering techniques or by exploiting software vulnerabilities.
In a decentralized architecture (P2P-based) there is no central point for communication. In this format of
structure the bots can act as either client or server or both [5]. Each bot is connected to some other bots of the
botnet. In this architecture botmaster can inject commands to a bot and afterwards this bot can broadcast it to
all other nodes in the network. Since decentralized botnet allows commands to be injected at any node in the
network, authentication of commands become essential to prevent other nodes from injecting incorrect
commands. So P2P- based botnet model impose a bigger challenge for defense of network [7]. Table.1 shows
different classifications of botnet based on their topological structure and some of their comparison
parameters.
TABLE I. C OMPARISON OF COMMAND AND CONTROL TOPOLOGIES
Design
Complexity

Delectability

Centralized

Low

Medium

Low

Low

Decentralized

Medium

Low

Medium

Medium

Unstructured

Low

High

High

High

Topology

Message
Latency

Survivability

This paper is presented a decentralized botnet detection model which is based on in-out degree of a node,
clustering and stability of a flow. The output of one module is input for the next module. In-out degree of a
node is useful to identify P2P node from a network. Because now adays attackers used decentralized
architecture so it is necessary to identify P2P node. After this module, clustering algorithm is used which is
based on association degree and distance between a pair of node. Finally botnet detection procedure has been
used to determine bots in a network. Because all the bots have higher symmetricity in order to get command
and in response. So packets related to bots are more stable than normal client packets.
The rest of the paper is organized as follows: Section II provides related works in the field of botnets. Section
III describes proposed botnet detection model and section IV concludes our work.
II. RELATED WORK
A botnet is a collection of zombie computers called bots. A bot is a compromised end computer, which might
be infected by malwares such as Trojan horses, backdoor, spywares and worms [16]. Bots will allow the
compromised computer to be controlled by the remote attacker i.e. Botmaster in the network. The botmaster
can act as head so that he/she is responsible for controlling and also issuing commands to launch many
different illegal criminal activities. In 1993, Eggdrop the first Internet Relay Chat (IRC) bot was developed to
automate IRC management activities [17]. These kinds of tools attract the hackers to write IRC-based
malicious botnets. Thereafter, more advanced botnets such as HTTP botnets and P2P botnets [11] [18] were
continued to appear.
In [3] [6], the authors discussed the impact of botnet attacks to computer world. Also they discussed the
methods to create botnet, propagation and communication technique, and protocols being used. An excellent
botnet tracking tool named Honeypot is used to monitor and understand the behavior of the bot operations
[4].
Yukiko Sawaya et al., [8] presented a flow-based attacker detection method which detect attackers (Nuisance
attackers, Simple attackers and Obvious attackers) without predefined blacklist/white list hosts. In their work,
they collected and analyzed traffic flows related to the object port (open port) and decoy port (closed port).
Since they used flow-based approach, attacks that were injected in payload would not be identified.
Alirena Shahrestani et al., [9] proposed visualization techniques which enhance the visibility of network
traffic related to invariant bot behaviors and provide notification of existence of bots. The main advantage
of this system is that visualized information is easier to be processed and for user it is easy to gain useful
knowledge about bot existence in a network. Their technique will be used to visualize a limited set of
invariant bot behavior (like fast response time, small size command etc.

105

Hsiao et al., [12] proposed a method which applied flow correlation for grouping the same activities of same
bots and then scoring technique to identify normal IRC and abnormal IRC behaviors. The authors of [10]
proposed a P2P botnet detection framework which is based on association between common P2P networks
behaviors and host behaviors. This mechanism not only detects known P2P botnet with a high detection rate
but also some unknown P2P malwares. This method deals with some problems such as data encryption, route
selection, communication behaviors.
S. Zander et al., [13] Proposed P2P traffic identification based on the signature of the key packets. Their
approach identified P2P flows on real-time traffic management and monitoring and based on signature of the
key packets, flow attributes and extracting signature from P2P flows. This approach has Low computational
complexity but generates false alarms if bots traffic is encrypted.
Yousof Al Hammadi et al., [14] proposed a behavioral correlation method for detecting P2P bots. Detection
of P2P botnet has been done using Log files analysis and by correlating bot behavioral attributes. Though
their method reduces false alarm, threshold to detection malicious process is unidentified.
In [15], a method to detect botnet based on flow based approach, packet sampling, payload based and pattern
matching has been proposed. It is a logical chance for high speed network. But absence of payload has to be
perceived.
III. PROPOSED WORK
A. System Architecture of Proposed Framework
The proposed System architecture given in Figure 1 contains the five modules, namely, packet capturing,
filtering, P2P node detection, clustering, and botnet detection. In this architecture it is clear that the input of
the next stage is the output of the previous stage.

Figure1. P2P botnet detection model

The dataset about the network is analyzed such that the packets flow in the network is captured using
wireshark tool in the server machine by making the network interface card as a promiscuous mode and the
sniffing is carried out. This captures the flow of packet passes through the server machine. The packet
captured is stored as a PCAP file in the database. The dataset may have infected host details as well as
uninfected host details in the network. For capturing packets we used wireshark tool. Wireshark is a free and
open source packet analyzer. Wireshark has graphical front end and plus some integrating storing and
filtering option. Wireshark has rich features set e.g. deep inspection of hundreds of hundreds of protocol,
multi platform, read/write many different capture file format etc.
Filtering module is responsible for filtering out unrelated traffic flows. Additionally it is responsible to filter
out packets into either legitimate or suspected from the network so that only the packets, which we want to
analyze are available in the network. So this stage reduces the traffic workloads.
Since goal of proposed system is to identify p2p botnet (decentralized bots) in a network, the first step is to
identify p2p nodes in the network. i.e. after filter out legitimate packets, it is necessary to check how many
nodes in a network received packet or message from other nodes. If any node in a network received botnet
packet then according to the definition of P2P botnet, these bots broadcast command or message to the
nearest nodes in the network or they find out victim nodes in the network. So at any particular time the nodes
106

incoming degree or outgoing degree which we call it as in-out degree is more. By using these related
concepts, we implemented a new algorithm, called In-Out Degree Algorithm (IODA) which will find P2P
nodes in the network. The number of P2P-nodes in the network can be identified based on the in-out degree
of a node.
The next step is to cluster these P2P nodes. It is based on similar behavior or similar communication. The
clustering has been done to identify the malicious peer in the network. K-Means Clustering Algorithm
(KMCA) has been used for P2P-node clustering. In this algorithm the node at the center is taken as a median
and then the next nearest node will be grouped and clustered as one node. The clustering can be made by
means of behavior of the node. Once the behavior value of the node differs from the normal one in the
cluster, then the node will be identified as suspicious node. The clustering can be made for the similar
behavior of the node that has been clustered. Then the dissimilar node can be grouped as one cluster and
these nodes will be analyzed in order to detect whether the node is a malicious one or not.
Finally P2P botnet traces are identified. This module uses Botnet Stability Detection Algorithm (BSDA) to
detect p2p bots in a network. The input for BSDA algorithm is the output of p2p clustering algorithm. In
BSDA algorithm, stability refers to the whether flows (in the context of botnet detection model flow refers to
the packets which contains information about the control and command packets i.e. information about size of
packets, timing information, same protocols etc.) are stable or unstable. If flows are stable it implies p2p
nodes are exchanging control or command packets to other nodes and accordingly definition of p2p bots,
these nodes are consider as a p2p botnet nodes. On the other hand if flows are unstable it implies p2p nodes
are exchanging packets other than C&C packets and these are not a p2p botnet nodes. The detection of the
botnet can be made by analyzing the control flow stability of a certain port such as UDP port. We can extract
control flows from bot traffic by collecting all flows on the selected UDP port.
B. P2P Node Detection
The algorithm for P2P node detection is based on in-out degree of a node. The in-out degree of a node ‘s’ is
within the set s(i) the number of members connected in a set to the total number of members in the set which
satisfies the condition.
Calculation of In-Out Degree of a Node S:
In-out degree of a node S is represented as IO(S)
Step 1: In a given specified time, collect the subset of nodes. Suppose, in time duration T, we collect subset
of nodes (samples of nodes) m times.
Step 2: Now for each subset of nodes, calculate the number of connections formed by each node. In ‘m’
subsets, number of connection formed by node S is N(m). So total number of connections formed by a node
S in all subsets = N (1) +N (2) +------------+N(m).
Average number of connections formed by
a node S in all m subsets (avg (S)) is N(1)+N(2)+----------+N(m) / m.
Step 3: In-out degree of a node S, in ith subset is defined as the total number of connections of a node S in ith
subset to the average number of connection of a node S in all subsets i.e. In-out degree IO(S(i)) = N(i)/ avg
(S). For each subset, we calculate the In-out degree of a node S. Let these degrees be (IOS(1),IOS(2)----------------IOS(m)) corresponding to every subset.
Step 4: Maximum In-out degree of a node S is
IO{S(max)}=MAX(IO{S(1)},IO{S(2)}-------IO{S(m)}).
Step 5: IO{S(maxx)} =IO{S(max)} * x%.
IO{S(minn)} = IO{S(max)} * y%. Where x and y are
constant. The value for x is 0.9 and y is 0.4.
Step 6: Now we define two sets:Set (1) = { IO{S(i)} /IO {S(i)} > IO{S(maxx)}}.
Where NOM (1) = number of members of that set.
Set (2) = {IO{S(i)}/IO {S(i)} < IO{S(minn)}}.
Where NOM (2) = number of members of that set.
Step 7: The In-out degree of a node S
IO(S) = NOM (2) / NOM (1).
Algorithm for p2p node detection (IODA)
Input: Set of In-out degree of a
node s in all m subset = ( IOS(1), IOS(2)-----------------IOS(m));
Initialize:IO{S(max)},IO{S(maxx)}, IO{S(minn)} NOM (1), NOM (2).
Begin
For i=1 to m do
If IO{S(i)} > IO{S(maxx)}, then
NOM (1) ++;
End;
If IO{S(i)} < IO{S(minn)}, then
107

NOM (2) --;
End;
End;
IO(S) = NOM (2) / NOM (1).
If IO(S) > S (IO)
Then node S is P2P node.
End;
End;
The values of S(IO) are the statistics characteristics of IO degree of a node, which we have derived from the
sample nodes.
C. P2P Node Clustering
For P2P node clustering, we used k-means algorithm which is based on association degree and distance
between a pair of nodes. K-means clustering is a method of cluster analysis which aims to partition n
observations into k clusters, in which each cluster belongs to the nearest mean.
Association degree of a pair of nodes [N(1,2) ] is
AD(1,2) = K(1) * [N(1,2)]+N(2,1)] + K(2) * T.
Where N(1,2) = number of packages that node 1 sent to node 2. N(2,1) = number of packages that node 2
sent to node 1. T= time of the connection between a pair of nodes (1,2) in the same period. K(1) and K(2) =
constant. Harmonious degree of a pair of nodes [N(1,2)] is
HD(1,2) = [N(1,2)+ N(2,1)] /ABS [N(1,2)
- N(2,1)].
The distance between a pair of nodes [N(1,2)] is
D(1,2) = 1 / [CD(1,2)].
The basic steps of KMCA algorithm are:
1.

Specify k points as the initial centroids.

2.

Repeat

3.

Form k clusters by assigning all points to the closet centroid.

4.

Recomputed the centroid of each cluster.

5. Until the centroids don’t change.
// specified k points as the initial centroids. //.
Step 1:- Select the initial-node of the C
categories (N1)1 , (N2)1 , (N3)1 ………. (NC)1. Let K=1.
// Repeat step 2, 3 and 4 till
(NJ)K+1 = (NJ)K //
Step 2:- Categorize all sampling nodes X into one specific category by K
times iteration.
If D[X(NJ)K] = min (1<J<K) D[X(NI)K ] Then X є (SJ)K, Where (SJ)K is the
K
set with center node (NJ) , where=1,2,3…………….C.
//recomputed the centroid of each
cluster. //
Step 3:- Use the category (SJ)K which obtained in step2 to update the
center node (NJ)K+1 of J category, until ∑J=1 to C ∑ X є (SJ)KD2 X(NJ)K+1 to be minimum value.
//
until the centroids don’t change. //
Step 4:- With all J=1,2,3…….C, if (NJ)K+1 = (NJ)K The
iteration ends, otherwise K = K+1, go step 2 and continue.
D. Botnet detection:
Finally botnet detection procedure which is based on stability of flow has been applied. The input for this
step is the output of the previous algorithm. A sequence of packets which contain same information i.e., same
source address same destination address, same source port, same destination port , same protocol etc., are
called flow. The flow record contains only address information, port information, timing information, size
information etc., but no information about payload. So for the detecting the stability of flows, we have to
select one information from flow packets. For this we select average number of bytes per packet / flow. For
simplicity, we denote this variable to avg(nbpp). For a flowj we divide the total number of bytes by the
number of packets within this flow and we can obtain the value of abp of this flow. In BSDA algorithm, we
use two timing windows and calculate the distribution of avg(nbpp) in different timing intervals and compare
these two distributions.
The Botnet Stability Detection Algorithm (BSDA):
Step 1:-specify each value of timing window one by one.
Steps 2:- specify the value of timing interval between these two timing windows.
Steps 3:- specify the value of flow change number
(initially zero) and epochs.
Steps 4:- calculate the distribution of avg(nbpp) in these two timing windows and compare.

108

Steps 5:- if they are closely related (difference is less than some threshold value which we calculated from
previous experiments) then print “flow is stable”.
Step 6:- After first interval, the second window (second avg(nbpp)) become the baseline for second interval
and again compute avg(nbpp) and compare.
Step 7;- if they are not closely related, report a flow change and increment the value of flow change number
by one and repeat the same procedure.
Step 8:- Repeat this procedure till flow not change.The timing interval is same between two avg(nbpp). So by
using this procedure we also able to say how many times flows changed.
IV. CONCLUSION
In this paper we have presented a P2P botnet detection model which is based on in-out degree of a node
(IODA), clustering (KMCA) and stability of flow (BSDA). In-out degree of a node is useful to identify P2P
nodes of a network. Clustering algorithm is based on harmonious degree and distance between a pair of
nodes. The botnet detection model is based on stability of flow. Since all bots follow the same rule and
predefined control and command method so by using this method we easily identify P2P bots from a
network. This proposed botnet detection model detects both known and unknown bots.

Figure 2. Relationship between avg number of connection and in-out degree of node

Figure 2 shows the relationship between average connection of nodes and in-out degree of nodes. Nodes
which generate more number of connections which we refer it as ‘avg’ number of connections with other
nodes having higher in-out degree than other node. i.e. they exchange more number of packets with other
nodes. In above graph average connections of node (in-out degree) 3 and 4 are high than all other nodes. So
this node is suspected as a P2P node while node 2 and 5 create less number of connections so these nodes are
considered to be legitimate P2P node, which is purely based on in-out degree of a node.

Figure 3. Relationship between association degree and harmonious degree of a pair of node

109

Figure 3 shows the relationship between a pair of nodes and their association degree and Harmonious degree.
By observing this graph it is clear that a pair of nodes which exchanges more number of packets, their
association degree and harmonious degree is higher than other nodes. This means that distance between them
is less than the other pairs of nodes. As inverse relationship exists between them, it tells about its
symmetricity too.

Figure 4. Stability index in different epochs

Finally Figure 4 shows how stability index vary according to control number and flow duration. As the
value of change number (CN) increase in different epochs, at the same time value of stability index (SI) also
increases in different epochs (keep timing interval same in all epochs). First value of CN is zero then value
of SI is one in all the epochs. As the value of CN slightly increases, there is more deviation in SI.

Figure 5. Matching number of bot features

The proposed work which has been implemented above using the specified standards and the previous work
is taken for experiment. The experimental result of both the work is evaluated by the parameter such as
matching number of bot features and stability index. The comparison graph is the evaluated results of both
the work and it is clearly shows that the proposed work has better matching number of bot features and
stability index compared to previous work. Figure 5 shows the matching number of bot features like attacks
port, SMTP error response etc. All bots were matched with any bot features, which eliminates the incidence
of any false negative on bot infected PCs. This implies that the implemented system can extract the bot
process and report in a stable manner.

110

REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]

[8]
[9]

[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]

Lei Zhang, Shui Yu, Di Wu, Paul Watters, “A Survey on Latest Botnet Attack and Defense”, International Joint Conference Of
IEEE Trustcom-11/IEEE ICESS-11/FCST-11, pp.53-60, 2011.
Maryam Feily, Alireza Shahreshtani and Sureswaran Ramadas “A Survey of Botnet and Botnet Detection”,IEEE in The Third
Conference of Emerging Security Information, Systems and Technologies, pp.79-84, 2009.
Banday, M.T. Qadri, J.A.Shah, “Study of Botnet and Their Threats to internet Security”, Sprouts:Working papers on Information
System, 2009. Available at:http://sprouts.aisnet.org/9-24.
Niels Provos, Thorsten Holz, “Virtual Honeypots: From Botnet Tracking to Intrusion Detection”, See Chapter11- Tracking
Botnets. Publisher: Addison Wesley Professional, Publishing 16 June, 2007.
Hossein Rouhani Zeidanloo, Farhoud Hossswinpour and Farhood Farid Etemad “New Approach For Detection of IRC and P2P
Botnet” International Journals of Computer and Electrical Engineering, Vol.2, No.6, pp.1029-1038, December, 2010.
Jing Liu, Yang Xiao, Kaveh Ghaboosi, Julia Deng, Jingyuan Zhang, “Botnet: Classification, Attacks, Detection, Tracing, and
Preventive Mesaures”, EURASIP Journal on Wireless Communications and Networking, 2008.
Mukesh Kumar, Pothula Sujatha, P. Manikandan, Madarapu, Naresh Kumar, Chetana Sidige and Sunil Kumar Verma “SelfDestructible Concentrated P2P Botnet” IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No.1, pp.401-405,
2011.
Yukiko Sawaya, Ayumu Kubota, Yutaka Miyake “Detection of Attackers in Services Using Anomalous Host Behavior Based on
Traffic Flow Statistics” PSJ International Symposium On Applications and the Internet, pp. 353-359,2011.
Alireza Shahrestani, Maryam Feily Rodina Ahmad, Sureswaran Ramadass “Discovery of Invariant Bot Behavior Through Visual
Network Monitoring System” Fourth International Conference on Emerging Security Information, Systems and
Technologies,pp.182-188, 2010
Yousof Al Hammadi, Uwe Aickelin “Behavioral Correlation for Detecting P2P Bots” Second International Conference on Future
Networks, DOI 10. 1109/ICFN 2010.72, pp. 323-327, 2010.
Hossein Rouhani Zeidanloo, Azizah bt Abdul Manaf “Botnet Detection by Monitoring Similar Communications Patterns”
International Journals of Computer Science and Information Security, Vol. 7,No.3, pp.36-44, 2010.
Hsiao-Chung Lin, Chia-Mei Chen, Jui-Yu Tzeng “Flow Based Botnet Detection” IEEE Fourth International Conference on
Innovative Computing, Information and control pp.1538-1541, 2009.
S. Zander, T. Nguyen and G. Armitage “P2P Traffic Identification Based on The Signature of Key Packets”, IEEE conference
Local Computer Networks, 2010.
Yousof Al Hammadi, Uwe Aickelin “Behavioral Correlation for Detecting P2P Bots” Second International Conference on Future
Networks, DOI 10. 1109/ICFN 2010.72, pp. 323-327, 2010.
Mohammed Abdul Qadeer, Mohammad Zahid “Network Traffic Analysis and Intrusion Detection Using Packet Sniffer” Second
International Conference on Communication Softwaret and Networks pp.313-317, 2010.
Tang T.K. and Wang J.H., “The Detection of zombies, Chinese cryptography and Information Security Association (CCISA 07)”.
Vol.13, July 2007, pp 25-36, ISSN 17729-6056.
Yang C.H. and Ting K.L.,” Fast Deployment of Botnet Detection with Traffic Monitoring”. Z. Fifth International Conference on
Intelligent Information Hiding and Multimedia signal processing (IIHMSP 09), Sep 2009, pp 856-860.
Lee J.S. Jeong, H.C., Park, J.H. Kim, M and Noh, B.N., “The Activity Analysis of Malicious HTTP-based Botnets using degree of
periodic Repeatability”, International conference on Security technology (SecTech 08), IEEE press, Mar 2008, pp 83-86.
DOI:10.1109/SecTech.2008.52.

111

A Dynamic Botnet Detection Model based on Behavior Analysis

More Related Content

What's hot (19)

Similar to A Dynamic Botnet Detection Model based on Behavior Analysis (20)

More from idescitation (20)

Recently uploaded (20)

A Dynamic Botnet Detection Model based on Behavior Analysis