Date of Award

2019

Degree Type

Thesis

Degree Name

Doctor of Philosophy

School

School of Science

First Advisor

Associate Professor Mike Johnstone

Second Advisor

Professor Craig Valli

Third Advisor

Dr Peter Hannay

Field of Research Code

08999

Abstract

Building Automation Systems (BAS) are a collection of devices and software which manage the operation of building services. The BAS market is expected to be a $19.25 billion USD industry by 2023, as a core feature of both the Internet of Things and Smart City technologies. However, securing these systems from cyber security threats is an emerging research area. Since initial deployment, BAS have evolved from isolated standalone networks to heterogeneous, interconnected networks allowing external connectivity through the Internet. The most prominent BAS protocol is BACnet/IP, which is estimated to hold 54.6% of world market share. BACnet/IP security features are often not implemented in BAS deployments, leaving systems unprotected against known network threats. This research investigated methods of detecting anomalous network traffic in BACnet/IP managed BAS in an effort to combat threats posed to these systems.

This research explored the threats facing BACnet/IP devices, through analysis of Internet accessible BACnet devices, vendor-defined device specifications, investigation of the BACnet specification, and known network attacks identified in the surrounding literature. The collected data were used to construct a threat matrix, which was applied to models of BACnet devices to evaluate potential exposure. Further, two potential unknown vulnerabilities were identified and explored using state modelling and device simulation.

A simulation environment and attack framework were constructed to generate both normal and malicious network traffic to explore the application of machine learning algorithms to identify both known and unknown network anomalies. To identify network patterns between the generated normal and malicious network traffic, unsupervised clustering, graph analysis with an unsupervised community detection algorithm, and time series analysis were used. The explored methods identified distinguishable network patterns for frequency-based known network attacks when compared to normal network traffic. However, as stand-alone methods for anomaly detection, these methods were found insufficient. Subsequently, Artificial Neural Networks and Hidden Markov Models were explored and found capable of detecting known network attacks. Further, Hidden Markov Models were also capable of detecting unknown network attacks in the generated datasets.

The classification accuracy of the Hidden Markov Models was evaluated using the Matthews Correlation Coefficient which accounts for imbalanced class sizes and assess both positive and negative classification ability for deriving its metric. The Hidden Markov Models were found capable of repeatedly detecting both known and unknown BACnet/IP attacks with True Positive Rates greater than 0.99 and Matthews Correlation Coefficients greater than 0.8 for five of six evaluated hosts.

This research identified and evaluated a range of methods capable of identifying anomalies in simulated BACnet/IP network traffic. Further, this research found that Hidden Markov Models were accurate at classifying both known and unknown attacks in the evaluated BACnet/IP managed BAS network.

Share

 
COinS