Vous êtes sur la page 1sur 52

NETWORK-BASED DETECTION OF IOT

BOTNET ATTACKS USING DEEP


AUTOENCODERS
By Aiswarya Ashok Kumar
S7 IT
Roll No: 5
Guide – Ms. Chinchu Krishna

1
OVERVIEW
■ Scope and objective
■ Introduction
■ Literature Review – Botnets and Related Work
■ Methodologies – Autoencoders and Main stages
■ Empirical Evaluation
■ Discussion and Results
■ Conclusion
■ References

2
SCOPE AND OBJECTIVE

■ A network based anomaly detection method


which uses deep autoencoders to detect
anomalous network traffic emanating from
compromised IoT devices.

3
SCOPE AND OBJECTIVE

■ Instantaneous detection can promote network


security, as it expedites the alerting and
disconnection of compromised IoT devices
from the network

4
INTRODUCTION

Number of IoT devices deployed are increasing


worldwide

Traffic volume of IoT-based DDoS attacks reaches


unprecedented levels

Need for timely detection of IoT botnet attacks has


become very crucial

5
WHAT ARE BOTNETS ?

■ Logical
connection of
Internet-
connected devices
■ Eg. Computers,
smartphones, IoT
devices
■ Security is
breached and
control ceded to a
third-party
6
BOTNET INFECTIONS

• Deploy botnets through a trojan horse


BOTHERDERS virus

USERS • Infect their own systems

•Access and modify personal information


BOTNET •Attack other computers and commit crimes
7
BOTNET FEATURES

■ Self – propagate : seek-and-infect missions


■ Search the web for vulnerable internet-connected
devices
lacking OS updates
antivirus

■ Botnet design continues to evolve


8
BOTNET FEATURES

■ Difficult to detect – users are anaware


■ Use only small amount of computing power
■ Botnets take time to grow – lay dormant

9
BOTNET ARCHITECTURE

10
BOTNETS - OPERATIONAL STEPS

■ Propagation
■ Infection
■ C&C Communication
■ Execution of Attacks

11
BOTNETS - OPERATIONAL STEPS

12
RELATED WORK – BOTNET DETECTION
METHODS

Specific operational step to


be detected

The detection approach

13
RELATED WORK

14
RELATED WORK

■ Previous studies focus on early operational steps


■ Botnet attacks mutate on a daily basis
■ Attacks become increasingly sophisticated
■ These mutations will eventually bypass existing
methods of early detection

15
RELATED WORK

■ Detection Approaches
 Host-based
 Network-based

16
RELATED WORK

■ Host – based : less realistic


 Installing detectors on IoT products

 Limited access to some IoT devices

 Constrained computation and power

 A single non-distributed solution is preferred

17
RELATED WORK

■ Network - based : uses deep learning to


perform anomaly detection
 Capture behavioral snapshot of benign IoT traffic
 Train a deep autoencoder – one for each device
 Autoencoder compress snapshots
• Failure to reconstruct snapshot - anomaly

18
RELATED WORK

■ Benefits of using network-based approach:

Heterogeneity Tolerance

Open World

Efficiency

Hardly any false alarms

19
METHODOLOGY

New data
•For each device •Detects anomalies
•Train on benign •Autoencoder •Device
traffic applied to new compromised
data of an IoT
device
Deep •Possibly infected
Anomalies
autoencoders

20
AUTOENCODERS

■ Unsupervised ANN
■ Learns to compress and encode data
■ Learns to reconstruct data back from coded
representation to a representation, as close as
to the original input

21
AUTOENCODERS

22
AUTOENCODERS
■ Reduces data dimensions – ignore noise in
data
■ Components:
 Encoder
 Bottleneck
 Decoder
 Reconstruction loss
■ Back propagation – reduce reconstruction loss
23
AUTOENCODERS

24
AUTOENCODERS FOR ANOMALY
DETECTION
■ Requires co-related input data
■ Encoding depends on co-related features
■ Train a autoencoder on a data set
Pass an image from that data set- reconstruction error
is low
Pass a random image/anomaly – reconstruction error
is high
25
AUTOENCODERS FOR ANOMALY
DETECTION

26
AUTOENCODERS FOR ANOMALY
DETECTION

27
AUTOENCODERS FOR ANOMALY
DETECTION

28
PROPOSED DETECTION METHOD

■ Concentrate on large enterprises with large


number of IoT devices connecting to their
networks via Wi-Fi
■ Devices include : self-deployed (e.g. Smart
smoke detectors) or dynamically introduced
(e.g. BYO Wearables)

29
PROPOSED DETECTION METHOD

■ Four main stages:


 Data Collection
 Feature Extraction
 Training an Anomaly Detector
 Continuous Monitoring

30
PROPOSED DETECTION METHOD

■ Data Collection:
 Capture raw traffic data
 pcap format
 Using port mirroring on the switch
 Collected immediately following installation

31
PROPOSED DETECTION METHOD

■ Feature Extraction:
Packet arrives – behavioral snapshot of hosts and
protocols
Obtain packet’s context by extracting 115 traffic
statistics over several temporal windows

32
PROPOSED DETECTION METHOD

 Summarize traffic that has:


Originated from same IP

Originated from both same source MAC and same IP address

Sent between the channel

Sent between the source to destination TCP/UDP sockets

33
PROPOSED DETECTION METHOD
 Extract same set of 23 features from five time windows of
the most recent 100ms, 500ms, 1.5 sec, 10sec, and 1min

34
PROPOSED DETECTION METHOD

■ Training an anomaly detector:


 Autoencoder trained to reconstruct its inputs after some
compression
 Compression ensures network learns meaningful
concepts
 Autoencoder fails at reconstructing abnormal
observations
 Significant reconstruction error - anomaly
35
PROPOSED DETECTION METHOD
 Optimize parameters and hyper parameters – maximize
TPR (True Positive Rate) and minimize FPR (False Positive
Rate)
 Training and optimization – two separate datasets, with
only benign data
 Training dataset DStrn – 2 input parameters:
 Learning rate, 𝛈

 No. of epochs – complete passes through entire DStrn


36
PROPOSED DETECTION METHOD

 Optimization dataset DSopt


 optimize 𝛈 and epochs iteratively until MSE between model’s
input and output stops decreasing

 Dsopt used to optimize a threshold (tr) which


discriminates between benign and malicious observations
 Anomaly threshold tr* = 𝑀𝑆𝐸 Ds opt + s(MSEDS opt )

37
PROPOSED DETECTION METHOD

 Single instance – high TPR & FPR


 To reduce FPR – base the abnormality decision on
a sequence of instances
 Determine minimal window size ws* as the
shortest sequence of instances – 0% FPR on DSopt

38
PROPOSED DETECTION METHOD

■ Continuous monitoring:
 Apply optimized model to vectors extracted from
continuously observed packets to mark as benign or
anomalous
 Majority vote on a sequence (length of ws*) of marked
instances – detect if entire stream is benign or anomalous
 Issue alert on detection of anomalous stream
39
EMPIRICAL EVALUATION

Mirai BASHLITE
■ Lab setup –
Replicate
organizational
data flow
■ Collect traffic
data from 9
IoT devices

40
EMPIRICAL EVALUATION

■ Botnets deployed
 BASHLITE – set a C&C server

 Mirai – C&C server + scanner + loader

41
EMPIRICAL EVALUATION
■ Attacks executed
 BASHLITE attacks
1) Scan: Scanning the network for vulnerable device
2) Junk: Sending spam data
3) UDP: UDP flooding
4) TCP: TCP flooding
5) COMBO: Sending spam data and opening a
connection to a specified IP address and port

42
EMPIRICAL EVALUATION

 Mirai Attacks
1) Scan: Automatic scanning for vulnerable devices
2) Ack: Ack flooding
3) Syn: Syn flooding
4) UDP: UDP flooding
5) UDP plain: UDP flooding with fewer options, optimized for higher
PPS

43
RESULTS AND DISCUSSION

■ Each of the nine sets of benign data from nine


IoT devices divided into 3 datasets:
 DStrn – training autoencoder
 Dsopt - optimizing parameters
 Benign part of DStst for estimating FPR

44
RESULTS AND DISCUSSION

■ Incorporate traffic from entire life cycle of


devices
■ Training and optimization use Keras
■ Each autoencoder input layer dimension
equals no. of features in dataset (115)

45
RESULTS AND DISCUSSION

46
RESULTS AND DISCUSSION

■ Same benign data to train three other


algorithms – Local Outlier Factor (LOF), One-class SVM
and Isolation forest

■ Optimize parameters and hyper parameters


■ Execute attacks

47
RESULTS AND DISCUSSION

48
CONCLUSION
■ Autoencoders for most IoT devices in a test set
obtained a zero FPR
■ Difficulty in capturing normal traffic behavior
varies among IoT devices and maybe
correlated with
 Device’s capabilities
 Network communications it normally produces
■ A solid predictability score can be leveraged by
large organizations 49
REFERENCES
■ N-BaIoT : Network-based Detection of IoT Botnet Attacks Using Deep
Autoencoders
Yair Meidan, Michael Bohadana, Yael Mathov, Yisroel Mirsky, Dominik Breitenbacher,
Asaf Shabtai, and Yuval Elovici
IEEE PERVASIVE COMPUTING, VOL. 13, NO. 9, JULY-SEPTEMBER 2018
■ Kitsune: An Ensemble of Autoencoders for Online Network Intrusion
Detection
Yisroel Mirsky, Tomer Doitshman, Yuval Elovici and Asaf Shabtai Ben-Gurion University of the
Negev
arXiv:1802.09089v2 [cs.CR] 27 May 2018

50
REFERENCES

■ Botnets and Internet of Things Security


E. Bertino and N. Islam
Computer, 2017
■ https://www.pandasecurity.com/mediacenter/security/what-is-a-botnet/
Date accessed : 4th October 2019
■ https://towardsdatascience.com/auto-encoder-what-is-it-and-what-is-it-used-for-
part-1-3e5c6f017726
Date accessed : 6th October 2019

51
THANK YOU

■ANY QUESTIONS ?

52

Vous aimerez peut-être aussi