Secure Multi-Satellite Communications in LEO Networks via Multi-Agent Deep Reinforcement Learning

In this paper, we propose a multi-agent deep reinforcement learning (MADRL) strategy for adaptive beamforming and artificial noise (AN) transmission to enhance physical layer security in low Earth orbit (LEO) satellite networks. Multiple satellites are jointly scheduled to cooperatively transmit data and AN against potential eavesdroppers such as hostile unmanned aerial vehicles. In the proposed scheme, each satellite independently selects its transmission mode (idle, data, or AN) and the corresponding beamforming vector to maximize the secrecy rate within a centralized training decentralized execution (CTDE) framework using the soft actor-critic (SAC) algorithm. The MADRL agents are trained using only statistical channel information of the adversary instead of full instantaneous channel state information. Simulation results demonstrate that the proposed scheme achieves a higher secrecy rate than conventional baseline schemes.