Eavesdropping defense scheme in C-V2X using deep learning and reinforcement learning

The broadcast nature of wireless signals allows malicious eavesdroppers to stealthily intercept and disrupt communication links, while dynamic spectrum competition in dense scenarios further degrades the performance of legitimate links. This paper investigates a collaborative security framework that integrates physical-layer dynamic authentication and deep reinforcement learning (DRL) for vehicular networks with multiple eavesdropping vehicles. To mitigate spoofing risks, we design a Long Short-Term Memory (LSTM)-based authentication model that verifies legitimate vehicle identities by extracting time-varying channel characteristics, thereby reducing communication risks with unauthorized vehicles. Upon ensuring legitimate identities, we propose a joint optimization scheme for transmit power control and channel selection. To address dynamic channel fading and high vehicle mobility, a Double Deep Q-Learning (DDQN) algorithm is developed to adaptively allocate limited spectral resources for maximizing secrecy capacity. Simulation results demonstrate that our LSTM-DDQN anti-eavesdropping model outperforms the related methods in system secrecy rate and V2V energy efficiency.