Abstract

Cyber-Physical Systems (CPS) are becoming more susceptible to sophisticated attacks that take
advantage of both physical and cyber components to degrade system operation while remaining
stealthy. Two significant contributions to the topic of CPS security are presented in this thesis.
First, a brand-new Data-Driven Covert-Replay Attack (DDCRA) is introduced, which uses recorded
input-output data to control the system while using covert signal cancellation to avoid detection.
The DDCRA is extremely useful and practical in real-world situations because it doesn’t require
knowledge of the system model or estimation technique, in contrast to conventional covert attacks.
Second, a dual-agent reinforcement learning detection architecture is presented to counter such
attacks. It consists of a Deep Q-Network (DQN) agent at the command and control center and a
Deep Deterministic Policy Gradient (DDPG) agent at the plant side. The DQN agent completes the
final attack detection after receiving anomaly signals from the DDPG agent, which keeps an eye on
system behavior locally. Simulations on a quadruple-tank process are used to verify the effectiveness
of the suggested assault and detection system. The results show that while the suggested RL-based
detector greatly increases detection accuracy in previous stealthy scenarios, the DDCRA can evade
detection from conventional residual-based techniques. This study enhances the state-of-the-art
in the area of intelligent CPS security mechanisms and emphasizes the necessity for more hybrid
solutions in the future.