Resilient path tracking control of cloud-based intelligent connected vehicle under hybrid cyber attacks: a safety-guided reinforcement learning approach.

Cloud-based intelligent connected vehicles (CICVs) are vulnerable to denial of service (DoS) and false data injection (FDI) attacks through wireless control channels, which can significantly degrade trajectory tracking performance and compromise driving safety. To improve control resilience and safety under hybrid cyber attacks, this paper proposes a safety-guided reinforcement learning control framework for CICVs. A unified discrete-time hybrid attack model is established to characterize the temporal, frequency, and amplitude constraints of hybrid cyber attacks, and reachable and controllable set analyses are performed to quantify the resulting degradation of system stability and safety margins. On this basis, a robust control Lyapunov function and control barrier function (CLBF)-based safety controller is developed by explicitly incorporating attack bounds into tightened stability and safety constraints. Furthermore, the tightened CLBF mechanism is embedded into the reinforcement learning framework through expert-guided policy learning, safety-aware reward shaping, and online action projection, thereby combining model-based safety assurance with data-driven performance optimization. Finally, simulations and cloud-based hardware-in-the-loop (C-HiL) experiments are conducted to validate the proposed method. The results show that, compared with the CLBF-based controller, the proposed method reduces the average lateral displacement error and average heading angle error by 87.42% and 94.57%, respectively, while achieving smoother control behavior and a larger safe controllable region under hybrid cyber attacks.