Advanced task-weighted meta-reinforcement learning for rapid secure routing in mobile RIS-assisted networks

This paper proposes a Reptile-based Meta-Reinforcement Learning (Meta-RL) framework for secure, adaptive multi-hop routing in Mobile Ad Hoc Networks (MANETs) enhanced by Reconfigurable Intelligent Surfaces (RIS). Addressing the retraining limitations of traditional Deep Reinforcement Learning (DRL), the framework employs task-weighted Reptile updates and a task-conditioned Markov Decision Process (MDP) to enable efficient adaptation to dynamic topologies and evolving threats. By integrating a stochastic secrecy constraint and quantized RIS control, it optimizes secrecy rate, latency, and energy efficiency under partial Channel State Information (CSI) uncertainty (σ ₑ²=0.05). Extensive simulations across 500 unseen tasks demonstrate superior performance over DRL, Model Agnostic Meta- Learning (MAML)-based Meta-RL, and Static Secure Path (SSP), achieving a 41% higher secrecy rate (2.5 vs. 1.77 bit/s/Hz), 51% lower Insecure Secrecy Rate (ISR) (0.1 vs. 0.204 bit/s/Hz), 18% reduced energy consumption (0.8 vs. 0.98 J/packet), 22% lower latency (12 vs. 15.4 ms), and 21% improved Packet Delivery Ratio (PDR) compared to MAML. The proposed Adaptation Efficiency Index (AEI) validates these gains, with Meta-RL converging in 450 episodes vs. 780 for MAML, ensuring robust secrecy and stability. Despite its lightweight, decentralized design, the framework faces high computational demands during offline meta-training. These advancements position it as a promising solution for next-generation wireless networks.