Learning-Based Vulnerability Analysis of Cyber-Physical Systems

This work focuses on the use of deep learning for vulnerability analysis of cyber-physical systems (CPS). Specifically, we consider a control architecture widely used in CPS, where the low-level control is based on a feedback controller and an observer (e.g., the extended Kalman filter (EKF)), while also employing an anomaly detector. To facilitate analyzing the impact potential sensing attacks could have on systems with general nonlinear dynamics, we develop learning-enabled attack generators capable of designing stealthy attacks that maximally degrade system operation. We show how such problem can be cast within a learning-based grey-box framework where only parts of the runtime information are known to the attacker. We then introduce two methods for generating effective stealthy attacks, based on feed-forward neural networks (FNN) and recurrent neural networks (RNN). Both types of attack-generator models are trained offline, using a cost function that combines the attack impact on the estimation error (and thus control) and the residual signal used for anomaly detection; this enables the trained models to recursively generate effective yet stealthy sensor attacks in real-time while requiring different levels of system information at runtime. The effectiveness of the proposed methods is demonstrated on several case studies with varying levels of complexity and nonlinearity: inverted pendulum, autonomous driving vehicles (ADV), and unmanned areal vehicles (UAVs).