Reinforcement Learning-Based Secrecy Capacity Maximization for RIS-Assisted Hybrid FSO/RF SAGIN

Ensuring physical-layer secrecy in Space-Air-Ground Integrated Networks (SAGIN) is challenging due to the broadcast nature of wireless links, volatile hybrid Free-Space Optical (FSO)/Radio Frequency (RF) channels, and the presence of eavesdroppers. This work addresses the problem of maximizing secrecy capacity in a RIS-assisted hybrid FSO/RF SAGIN, where a satellite communicates with ground users through a High-Altitude Platform (HAP) relay and UAV-mounted reconfigurable reflecting platforms. The joint optimization of transmit power, mode-dependent reflecting configuration, and adaptive FSO/RF mode selection presents a highly non-convex control problem. To overcome these challenges, we reformulate the task as a Markov Decision Process (MDP) and design a Double-Dueling Deep Q-Network (Double-DDQN) with Prioritized Experience Replay (PER-DDQN). The agent learns power levels, mode-dependent reflecting codewords, and link-selection actions in real time using environment-aware reward shaping. Experimental results demonstrate that the proposed PER-DDQN significantly outperforms Vanilla DQN, Plain DQN, and uniform replay in secrecy capacity, convergence speed, and training stability. Furthermore, reflecting-platform ablation studies confirm that the proposed UAV-mounted reconfigurable platform, especially with larger apertures, substantially enhances secrecy while maintaining favorable secrecy-energy efficiency.