Multi-Agent Learning for Cooperative Anti-Jamming Data Collections in UAV-Assisted Wireless Networks

Abstract

In this paper, we focus on anti-jamming communications in an uncrewed aerial vehicle (UAV)-assisted wireless network comprising multiple UAVs and a mobile UAV jammer (MUJ). The MUJ can observe the UAVs' locations and adaptively update its jamming strategy by analyzing the detected spatial distribution of the UAV swarm to identify the currently active collector UAVs. In response to such threats, we consider a deception-enabled strategy in which each UAV can dynamically switch between collector mode and decoy mode. Collector UAVs gather data from ground users (GUs), whereas decoy UAVs deliberately maintain spatial proximity to lure the adversarial jamming beam away from the collectors. The interaction between the decoy UAVs and the MUJ's adaptive beamforming constitutes a complex spatio-temporal game, which introduces a trade-off between deception effectiveness and data collection efficiency. To address this trade-off, we formulate a throughput maximization problem by jointly optimizing the UAVs' trajectories and mode switching strategies. Given the highly dynamic and non-stationary nature of this adversarial problem, we propose a multi-agent deep reinforcement learning (MADRL) approach that integrates a deception cost to balance jamming suppression gains and data throughput. Simulation results demonstrate that the proposed framework outperforms baseline schemes in terms of anti-jamming throughput, validating the effectiveness of the decoy-based deception strategy and dynamic mode switching mechanism.