A Federated Distributed Spectrum Sharing Algorithm for UAV Swarm

To address the real-time and fault-tolerant requirements of dynamic spectrum resource allocation for unmanned aerial vehicle (UAV) swarm in time-varying environments, this paper proposes a clustered federated distributed spectrum-sharing algorithm based on federated learning. The algorithm constructs a dual-channel deep reinforcement learning network, employs an adaptive adjustment architecture, and achieves dynamic training mode switching through a two-phase collaborative mechanism. When the central UAV is functioning normally, a federated vertical reinforcement training algorithm based on independent dual-layer DuelingDQN is employed. In the event of central UAV failure, a federated horizontal training algorithm incorporating transfer learning and a voting mechanism is utilized. This approach achieves breakthroughs in three key aspects: (1) utilizing a clustered federated distributed spectrum learning framework to effectively suppress channel interference; (2) mitigating the risk of single-node failures through horizontal transfer techniques; and (3) significantly enhancing training convergence efficiency via a parallel training mechanism based on a dual-channel deep reinforcement learning network. Experimental results show that compared to traditional DQN and DuelingDQN algorithms, the proposed algorithm exhibits significant superiority in convergence speed, final model decision-making capability, and robustness.