Deep Reinforcement Learning for Semantic Secure Energy Efficiency Optimization in IRS-Assisted UAV Communications

The integration of semantic communication with unmanned aerial vehicles (UAVs) and intelligent reflecting surfaces (IRSs) offers a promising approach for next-generation wireless systems. However, jointly optimizing semantic reliability, physical-layer security, and energy efficiency remains challenging. This paper investigates an IRS-assisted UAV semantic secure communication system in the presence of a potential eavesdropper, with the objective of maximizing semantic secure energy efficiency (SSEE). We first introduce a semantic similarity-based secure energy efficiency metric to capture the trade-off among transmission reliability, physical-layer security, and UAV energy consumption. The semantic symbol number, UAV trajectory, transmit power, and IRS phase shifts are jointly considered under mobility, secrecy, and energy constraints. To efficiently solve the resulting mixed discrete-continuous optimization problem, we develop a hierarchical optimization framework: the outer layer exhaustively searches over a finite set of candidate semantic symbol numbers, while the inner layer solves the continuous resource allocation problem using a heuristic-guided soft actor-critic (HG-SAC) algorithm. Simulation results show that the proposed framework achieves noticeable SSEE gains over several benchmark schemes.