Abstract
This paper addresses the challenging problem of discrete phase configuration for reconfigurable intelligent surfaces (RISs) in secure non-terrestrial networks (NTNs). The combination of discrete hardware constraints and secrecy-driven objectives creates a highly non-convex and combinatorial optimization problem. To address this challenge, we propose a novel heuristic conditional generative adversarial network (cGAN) framework that enables efficient and secure RIS phase control through a data-driven approach. Our approach employs heuristic algorithms, such as greedy optimization, to generate high-quality training labels that guide the conditional learning process. To enable differentiable learning over discrete phase values, we integrate a Gumbel-Softmax-based quantization mechanism into the generator architecture. Additionally, we introduce a multi-sample inference and selection strategy that enhances generalization and stability by generating multiple candidate RIS configurations and selecting the optimal one based on secrecy rate maximization. Extensive simulation results demonstrate that the proposed framework achieves near-optimal secrecy rate performance while significantly reducing computational complexity and inference latency compared to exhaustive search and traditional greedy methods. The framework exhibits strong robustness across various channel conditions and quantization levels, confirming its practical viability for real-time applications. This work demonstrates the potential of integrating heuristic principles with generative learning for scalable and secure RIS configuration in dynamic NTN environments, thereby providing a foundation for future research in learning-based wireless optimization and intelligent network control.
Funding Information
  • Ericsson Canada
  • Natural Sciences and Engineering Research Council of Canada