An Optimal Hybrid Learning Approach for Attack Detection in Linear Networked Control Systems  

作　　者：Haifeng Niu Avimanyu Sahoo Chandreyee Bhowmick S·Jagannathan 

机构地区：[1]Seattle,WA 98103 USA [2]School of Electrical and Computer Engineering,Oklahoma State University,Still Water,OK 74078 USA [3]Department of Electrical and Computer Engineering,Missouri University of Science and Technology,Rolla,MO 65409 USA

出　　处：《IEEE/CAA Journal of Automatica Sinica》2019年第6期1404-1416,共13页自动化学报（英文版）

基　　金：supported in part by the National Science Foundation(IIP 1134721,ECCS 1406533,CMMI 1547042)

摘　　要：A novel learning-based attack detection and estimation scheme is proposed for linear networked control systems(NCS),wherein the attacks on the communication network in the feedback loop are expected to increase network induced delays and packet losses,thus changing the physical system dynamics.First,the network traffic flow is modeled as a linear system with uncertain state matrix and an optimal Q-learning based control scheme over finite-horizon is utilized to stabilize the flow.Next,an adaptive observer is proposed to generate the detection residual,which is subsequently used to determine the onset of an attack when it exceeds a predefined threshold,followed by an estimation scheme for the signal injected by the attacker.A stochastic linear system after incorporating network-induced random delays and packet losses is considered as the uncertain physical system dynamics.The attack detection scheme at the physical system uses the magnitude of the state vector to detect attacks both on the sensor and the actuator.The maximum tolerable delay that the physical system can tolerate due to networked induced delays and packet losses is also derived.Simulations have been performed to demonstrate the effectiveness of the proposed schemes.A novel learning-based attack detection and estimation scheme is proposed for linear networked control systems(NCS), wherein the attacks on the communication network in the feedback loop are expected to increase network induced delays and packet losses, thus changing the physical system dynamics.First, the network traffic flow is modeled as a linear system with uncertain state matrix and an optimal Q-learning based control scheme over finite-horizon is utilized to stabilize the flow. Next, an adaptive observer is proposed to generate the detection residual,which is subsequently used to determine the onset of an attack when it exceeds a predefined threshold, followed by an estimation scheme for the signal injected by the attacker. A stochastic linear system after incorporating network-induced random delays and packet losses is considered as the uncertain physical system dynamics. The attack detection scheme at the physical system uses the magnitude of the state vector to detect attacks both on the sensor and the actuator. The maximum tolerable delay that the physical system can tolerate due to networked induced delays and packet losses is also derived. Simulations have been performed to demonstrate the effectiveness of the proposed schemes.

关 键 词：ATTACK detection ATTACK estimation event-triggered CONTROL Lyapunov stability networked CONTROL system(NCS) OPTIMAL CONTROL Q-LEARNING 

分 类 号：TN9[电子电信—信息与通信工程]