Abstract

Cyber-physical systems (CPS) are integral to critical infrastructures such as power networks, transportation systems, and water treatment networks. Despite the advancements in developing more secure CPS and monitoring systems, the number of successfully executed cyber-attacks in CPS has increased over the past decade. The mentioned cyber-attacks, which can make CPS unstable, are performed by intelligent adversaries who try to maintain their malicious attacks undetected. This thesis addresses several crucial challenges related to cyber-attacks in CPS and multi-agent systems (MAS).

The first part of the thesis focuses on simultaneous cyber-attacks and fault detection and isolation (CAFDI) in centralized and large-scale interconnected CPS. Proposed methodologies include centralized and distributed CAFDI approaches, incorporating two filters and an unknown input observer (UIO)-based detector to identify various deception attacks such as covert, zero dynamics, and replay attacks. The effectiveness of the distributed CAFDI approach is demonstrated through a hardware-in-the-loop (HIL) simulation of a four-area power network system.

The second part studies stealthy cyber-attacks in CPS, particularly zero dynamics, covert, and controllable attacks. Conditions for executing these attacks are derived from CPS Markov parameters and the system observability matrix. Dynamic coding schemes are proposed as countermeasures, increasing the number of actuators needed to execute cyber-attacks.

In the third part, zero dynamics and undetectable cyber-attacks in linear and nonlinear CPS are explored. A new security metric, security effort (SE), is introduced to determine the minimum number of secured actuators and sensors required to prevent such attacks in linear CPS. For nonlinear CPS, the study uses Koopman operator theory and the extended dynamic mode decomposition (EDMD) algorithm to create a finite-dimensional linear representation of the system to identify critical sensor measurements that need securing to prevent zero dynamics and covert attacks.

The fourth part addresses privacy-preserving consensus control, controllability cyber-attacks, undetectable cyber-attacks, and detection methodologies in MAS. A distributed transformation-based consensus control method is developed to protect agent privacy from eavesdroppers. Conditions for adversaries to control the MAS network by attacking a few agents are explored, defining these as controllability cyber-attacks. Undetectable cyber-attacks in MAS are defined and an event-triggered detection module to detect such attacks is proposed.