Enhancing Physical Layer Security in Cell-Free Networks Through IRS-Assisted Meta-Reinforcement Learning

This paper provides a framework to enhance physical layer security (PLS) in cell-free massive multiple-input multiple-output systems (CF-mMIMO). This system combines multiple intelligent reflecting surfaces (IRSs) with a type of machine learning called meta-reinforcement learning (Meta-RL). The proposed method addresses the primary challenge of protecting distributed wireless transmissions from sophisticated eavesdropping, taking into account channel state information (CSI), hardware constraints, and quantized phase control. This research presents a joint optimization method achieved through the exploitation of active beamforming at distributed access points (APs) and passive beamforming at multiple IRS panels via a combined Meta-RL and Q-learning algorithm. The system adapts to changing channel conditions, estimation inaccuracies, and eavesdropping risks while maintaining computational efficiency with a decoupled two-timescale optimization structure. The system model explicitly incorporates CSI estimation errors, channel degradation over time, multiple IRS coordination, and authentic hardware limitations. Extensive simulations show that the suggested approach surpasses conventional benchmarks in terms of secrecy rate, energy efficiency, and speed of convergence. In particular, it achieves up to 45% higher secrecy rates and 38% better energy efficiency compared to the most effective current approaches. It remains reliable even under substantial CSI imperfections (error variance up to 0.5), discrete phase quantization (3 to 5 bits), and multi-IRS coordination overhead. Two-timescale stochastic approximation is used to provide theoretical convergence guarantees, with practical feasibility supported by a detailed analysis of complexity.