Deep Reinforcement Learning Aided Trajectory and Power Control for Secure UAV Communication

This study investigates a physical layer security-based Unmanned Aerial Vehicle(UAV) communication system aimed at protecting confidential signals from being intercepted by eavesdroppers. The system utilizes dual UAVs, where one UAV serves as an aerial base station for communication with ground users, and the other UAV acts as an interference device to deceive eavesdroppers. In the presence of multiple eavesdroppers and legitimate users, a deep reinforcement learning (DRL) approach called Deep Q-Networks(DQN) is proposed. It jointly optimizes the trajectory of the UAV, the transmission power of the UAV transmitter, and user scheduling to maximize the confidentiality performance in the presence of multiple eavesdroppers. For high-dimensional and continuous state space in the training process, DQN effectively improves its convergence performance in offline learning by implementing experiential replay and target network technology. Simulation results demonstrate that compared to the other two baseline methods, this approach achieves faster convergence speed and better secrecy rate performance.