Attack Detection in Linear Networked Control Systems by Using Learning Methodology

A novel learning-based attack detection scheme for linear networked control systems (NCS) is introduced. The class of attacks considered here tends to increase network induced delays and packet losses which affects the physical system dynamics. For the network side, an adaptive observer is proposed to generate the attack detection residual, which in turn is utilized to determine the onset of an attack when it exceeds a predefined threshold. The uncertain stochastic physical system dynamics as a result of network-induced delays of packet losses require an optimal Q-learning based event-triggered controller that optimizes the control policy and the event-triggering instants simultaneously. The magnitude of the state vector of the physical system works as the detection signal for both sensor and actuator attacks. The maximum tolerable delay by the physical system due to packet losses is also derived. Simulations are included to demonstrate the effectiveness of the proposed schemes using the network and sensor/actuator attacks.