Secure NOMA-Enabled Aerial RIS Networks via Multi-Agent Multi-Stage Curriculum Learning

The broadcast nature of wireless channels makes sixth-generation (6 G) networks vulnerable to eavesdropping, necessitating robust physical layer security (PLS). In this paper, we propose a secure non-orthogonal multiple access (NOMA)-enabled unmanned aerial vehicle (UAV)-mounted reconfigurable intelligent surface (RIS) architecture to maximize secrecy rates in the presence of eavesdroppers. UAV-mounted RIS enables dynamic repositioning to establish favorable line-of-sight (LoS) links while suppressing signal leakage. However, jointly optimizing UAV 3D trajectory, RIS phase shifts control, base station (BS) active beamforming, and power allocation is challenging due to tightly coupled variables, strict successive interference cancellation (SIC) constraints, and minimum secrecy requirements. We formulate a secrecy rate maximization problem as a highly non-convex mixed-integer nonlinear program (NC-MINLP), intractable for conventional methods. We reformulate it as a Markov decision process (MDP) and propose a multi-agent multi-stage curriculum learning framework with proximal policy optimization (MAMSCL-PPO). The framework progressively expands the secured user set, addressing reward sparsity and enabling knowledge transfer. Two cooperative agents, employing centralized training with decentralized execution (CTDE), jointly control the UAV trajectory with RIS phase shifts and the BS beamforming with power allocation. Simulations demonstrate that MAMSCL-PPO achieves a 43% higher sum secrecy rate and faster convergence than single-stage training, validating its effectiveness for scalable PLS in NOMA-enabled aerial networks.