Cooperative multi-agent DRL-based robust secure transmission in UAV-assisted HetNets

This paper investigates robust physical-layer security (PLS) for UAV-assisted double-layer heterogeneous networks (HetNets) under channel state information (CSI) uncertainty. A joint channel assignment and power control optimization problem is formulated to maximize system sum capacity while satisfying worst-case secrecy constraints. To address the resulting mixed-integer nonlinear programming (MINLP) problem and multi-agent non-stationarity, we propose a Robust Action-Masked Multi-Agent Deep Reinforcement Learning (RAM-MADRL) framework. In this framework, the action-masking mechanism enforces physical feasibility during decision making, while the cooperative mini-batch mechanism is adopted as a training strategy to improve learning stability under the CTDE architecture. Simulation results under a representative setup with four UAVs, seven U2A links, a secrecy-rate requirement of c 0 = 2 bps/Hz, and a UAV altitude of 20 m show that the proposed RAM-MADRL achieves about a 10% improvement in system sum capacity over learning-based baselines while maintaining robust secrecy performance under CSI uncertainty.