Deep Reinforcement Learning-Based Resource Allocation for Secure RIS-aided UAV Communication

We investigate the use of reconfigurable intelligent surfaces (RISs) in wireless networks to maximize the sum secrecy rate (i.e., the sum maximum rate that can be communicated under perfect secrecy). Specifically, we focus on a network that utilizes RIS-assisted unmanned aerial vehicles (UAVs) under imperfect channel state information (CSI). Our objective is to maximize the sum secrecy rate while dealing with the presence of multiple eavesdroppers. To achieve this, we jointly optimize the active (UAV) and passive (RIS) beamforming together with the UAV's trajectories. The formulated problem is non-convex due to the coupling of CSI with the maneuverability of the UAV. To overcome this challenge, we propose a policy-based deep reinforcement learning (DRL) approach that solves the non-convex optimization problem in a centralized fashion. Finally, simulation results show that our proposed approach significantly improves average sum secrecy rates over conventional approaches.