An IRL Approach for Cyber-Physical Attack Intention Prediction and Recovery

Abstract

Modern autonomous cyber-physical systems have been demonstrated to be vulnerable to cyber-attacks like sensor spoofing in which an attacker stealthily compromises sensor readings to hijack the system toward undesired and possibly hazardous states. The majority of security techniques developed today are, however, reactive and concerned with detection and interdiction of attacks without considering predicting the intention of the attack. To deal with such problem, we propose a Bayesian Inverse Reinforcement Learning technique that leverages the history of sensor data and control inputs to predict the goal of sensor spoofing attacks, determine which sensors are compromised, and recover the system. We also propose an active exploration framework to improve the convergence rate of the intention inference before reaching undesired states. The proposed approach is validated with simulation results on a quadrotor unmanned aerial vehicle navigation case study.