RIS-shield: intelligent reflecting surface-aided physical layer security with deep reinforcement learning for 6G massive IoT

This paper proposes RIS-Shield, an intelligent defense architecture for secure 6G massive Internet of Things (MIoT) deployments. By integrating Reconfigurable Intelligent Surface (RIS) technology with deep reinforcement learning (DRL), RIS-Shield addresses critical physical layer security (PLS) challenges that traditional mathematical optimization methods fail to solve under high-dimensional non-convex conditions. Specifically, we formulate secure resource allocation as a Partially Observable Markov Decision Process (POMDP) modeling and develop a lightweight decision framework that couples Graph Neural Networks (GNN) for spatial feature extraction, multi-head self-attention for dynamic threat prioritization, and meta-learning accelerated proximal policy optimization (PPO) for rapid policy adaptation. The proposed framework leverages graph-structured representations to capture spatial topology relationships among massive terminals, enabling efficient interference coordination and dynamic threat response. The GNN–attention–PPO pipeline compresses high-dimensional channel observations through sliding-window temporal convolution and topological graph abstraction, achieving sub-4-ms inference latency on resource-constrained edge platforms while maintaining near-optimal security performance relative to the alternating optimization with semidefinite relaxation (AO-SDR) theoretical upper bound. Experimental results, averaged over five independent random seeds and three distinct channel models, demonstrate that RIS-Shield achieves approximately 50% fewer steps than twin delayed deep deterministic policy gradient (TD3) and deep Q-network (DQN) baselines, with 98.2% of decisions meeting 6G ultra-reliable low-latency communication (URLLC) scheduling windows. The meta-learning-driven initialization further reduces cold-start convergence steps by 57% (standard deviation: 8.3%) across diverse channel distributions and dynamic topology configurations. The dynamic eavesdropper tracking and heterogeneous service multi-objective optimization experiment validate real-time adaptive capability against mobile threats and differentiated quality-of-service (QoS) guarantees for industrial sensing, video streaming, and environmental monitoring.