Secure Cooperative Communications in 6G Networks: A Constrained Hierarchical Reinforcement Learning Framework with Hybrid Action Space.

With the rapid evolution toward 6G networks, ensuring robust physical layer security (PLS) in highly dynamic and heterogeneous wireless environments has become a key challenge. Traditional security methods often struggle to adapt to time-varying channels, especially in the absence of perfect channel state information. Furthermore, the dynamic nature of node selection and power allocation in heterogeneous networks creates a complex hybrid action space operating across multiple timescales, significantly complicating the design of efficient and adaptive security strategies. To address this, this paper proposes a novel constrained hierarchical reinforcement learning (CHRL) framework for secure cooperative communications in next-generation wireless systems. The framework is designed to optimize secrecy performance within a hybrid action space comprising both discrete node selection and continuous power allocation, operating at different timescales. By hierarchically decoupling the joint optimization problem, the upper layer performs risk-aware node selection to maximize long-term secrecy capacity (SC) while guaranteeing a stable and secure link. At the lower layer, we develop a constrained MiniMax Multi-objective Deep Deterministic Policy Gradient (M3DDPG) algorithm that optimizes power allocation considering worst-case conditions. Lagrange multipliers are integrated to enforce a strictly positive SC constraint throughout transmission, effectively preventing security outages. Simulation results under time-varying Rayleigh fading channels demonstrate that the proposed CHRL framework outperforms existing HRL methods, achieving up to 17% improvement in SC while strictly maintaining security constraints. These results validate the effectiveness of the proposed approach for enhancing PLS in next-generation cooperative wireless networks.