Trust-Based Information Filtering for Robust Decentralized Execution of Pre-Trained MARL Policies in UAV Swarms

Multi-Agent Reinforcement Learning (MARL) enables complex drone swarm coordination; however, the mission success is hindered by unreliable communication. Existing corrective solutions often require integration during training or significant configuration, limiting flexibility. This paper introduces a Trust-Based Information Filtering (TIF) system that enhances pre-trained MARL policies during decentralized execution. The post-hoc TIF system equips each agent with a mechanism to assess message trustworthiness using learned spatiotemporal expectations from normal operations. This dynamic self-configuration eliminates the need for attack data or policy retraining. Evaluated in UAV formation control under various communication unreliability scenarios, TIF demonstrates a measurable improvement in operational resilience. This validates the prototype of effective, lightweight, post-hoc filtering approach, signaling that robustness can be layered onto existing MARL policies without costly retraining.