Lijing Zhai
(Advisor: Prof. Kyriakos G. Vamvoudakis]

will defend a doctoral thesis entitled,

Architectures for Hardening Security in Intelligent Cyber-physical Systems

On

Wednesday, May 24 at 11:00 a.m.
Montgomery Knight Building 317
[Click here to join the meeting]

Abstract
Cyber-physical systems, consisting of interconnected physical devices and computational and communication components, are susceptible to adversarial attacks that can compromise their functionality, as well as to uncertainties that may arise within their information-sharing architectures. The continued growth of the Internet of Things is driving the development of applications in which networked systems gather and collectively process data to enable better decision making. As such, a pressing concern is the security of these systems. To fully realize the potential of cyber-physical systems in this context, it is critical to consider the roles of decision-making mechanisms, information exchange architectures, security evaluation under adversarial conditions, and security guarantees in dealing with uncertainties while leveraging data.

In this thesis, we will present principles and methods that enable the efficient and resilient operation of autonomous systems in adversarial and uncertain environments by taking advantage of control and learning theories. Our work revolves around hardening the security of cyber-physical systems from three perspectives: data-driven defense mechanisms, security quantification in the presence of adversaries, as well as safety guarantees under uncertainties.

With abundant data generated by cyber-physical systems, the initial focus will be on developing defense mechanisms against adversarial attacks in a data-driven manner. This line of work will encompass both reactive measures, i.e., statistics-based defense against replay attacks, and proactive measures, i.e., unpredictability-based defense against attacks on actuating and sensing components. Taking advantage of redundant components, the unpredictable changes in system dynamics can diminish the effectiveness of adversary reconnaissance efforts. Subsequently, we will adopt a more abstract perspective to evaluate the security of cyber-physical systems in the presence of adversaries. We will concentrate on identifying the conditions under which malicious attacks go undetected. Graph theory will be employed to develop an undetectability-based security quantification and to investigate methods for efficient computation. This security quantification will provide valuable guidance for decision making concerning component placement and allocation in cyber-physical system design. Finally, in the context of uncertainties, we will delve into the resilience of cyber-physical systems with a particular emphasis on learning behaviors. To account for time asynchronization between various components, we will investigate the impact of clock offsets on data-driven reinforcement learning algorithms. In the stochastic setting, we will propose a policy iteration algorithm for stochastic systems subject to probabilistic constraints, providing stability guarantees.

Committee

• Prof. Kyriakos G. Vamvoudakis – School of Aerospace Engineering (advisor)
• Prof. Wassim M. Haddad – School of Aerospace Engineering
• Prof. Samuel D. Coogan – School of Electrical and Computer Engineering
• Prof. Tamer Başar – Department of Electrical and Computer Engineering, University of Illinois Urbana-Champaign