Prediction-enhanced deep reinforcement learning for RIS-aided high-speed railway security under unknown eavesdropping

As high-speed railway (HSR) services become increasingly data-intensive, secure train-to-ground transmission is essential for dependable onboard connectivity and operation-related information exchange. However, in reconfigurable intelligent surface (RIS)-aided HSR communications, physical-layer security optimization remains challenging when instantaneous eavesdropper (Eve) channel state information (CSI) is unavailable and legitimate CSI becomes outdated under high-speed mobility. To address this issue, this paper develops a prediction-enhanced deep reinforcement learning framework for RIS-aided HSR communications under unknown eavesdropping. First, a secrecy-rate lower bound is constructed from the statistical prior of the Eve link, enabling conservative security evaluation without instantaneous Eve CSI. The RIS reconfiguration cost is further embedded into the optimization objective, forming an effective security utility that balances secrecy enhancement and phase-switching overhead. Second, an attention-enhanced long short-term memory (Attention-LSTM) prediction module is designed to infer the current legitimate-channel state from historical channel observations and mitigate CSI aging. The predicted channel state and previous RIS phase are then organized as a prediction-enhanced state, based on which joint beamforming and RIS phase control is formulated as a continuous Markov decision process (MDP). A twin delayed deep deterministic policy gradient (TD3)-based agent is developed to learn the online control policy for adaptive beamforming and RIS phase optimization over time-varying railway channels. Simulation results show that the proposed framework improves legitimate-channel prediction accuracy, accelerates policy convergence, increases effective security utility, and reduces unnecessary RIS reconfiguration compared with representative baseline schemes.