Deep Reinforcement Learning based Secrecy Spectrum Reuse Optimization for C-V2X Communications

This paper examines a deep reinforcement learning (DRL) approach for optimizing secrecy spectrum reuse in cellular vehicle-to-everything ($\mathbf{C}-\mathbf{V 2 X}$) communications. The scenario includes multiple V 2 V clusters (V -clusters) and V 2 V pairs (V pairs) within a macro cell’s coverage, all sharing a wireless spectrum through time-division reuse. Macro cell users (MCeUs) and V-clusters employ distinct reuse patterns, while V-pairs must dynamically adjust their reuse strategies based on MCeUs’ locations to avoid severe interference, especially for MCeUs at the cell edge. This is challenging due to limited knowledge of other users’ reuse behaviors. DRL enables V-pairs to learn an effective spectrum reuse strategy without prior information. We propose a double deep Q-network based algorithm, termed D2S2R, which allows V-pairs to select appropriate time slots for spectrum reuse, maximizing the total throughput of all users. To ensure stable throughput for V-pairs, the algorithm integrates stability into its objective function. Simulations indicate that the D2S2R algorithm achieves near-optimal throughput, closely matching the benchmark (theoretical throughput boundary) across different cases, and significantly outperforms methods relying on base station cooperation.