Signal Reconstruction and Memory Feedback Control for Path Tracking under Penetrative Cyber-attacks

Networked control systems provide greater flexibility for autonomous vehicles but also increase vulnerability to cyber-attacks. In particular, penetrative injection attacks on intelligent sensors can corrupt valid data and jeopardize driving safety. To this end, a reliable path tracking control scheme is developed for autonomous networked vehicles, capable of maintaining performance even when some or all sensor signals are compromised. The proposed method integrates attack detection and signal reconstruction within a resilient switching control framework to improve robustness. Specifically, a deep physics-informed neural networks (PINN), trained on attack-free data and a physical model whose parameters are refined offline via an extended Kalman filter (EKF), is employed to model vehicle dynamics and reconstruct vehicle state signals. These estimates are subsequently refined and uncertainty-quantified using a Gaussian processes (GPs)-based model, enabling accurate signal reconstruction and adaptive attack detection. Depending on the identified attack status, the system dynamically switches between a conventional robust controller and a memory-based feedback controller. The memory-based controller leverages trusted historical data feedback and stability inference to maintain tracking performance under sensor attacks. Simulation and experimental results demonstrate that the proposed approach improves lateral and heading accuracy by 83.93% and 78.54%, respectively, and enhances vehicle stability by 80.47%, compared with methods that do not account for cyber-attacks.