Zero-trust graph-attentive intelligence for Byzantine-resilient UAV swarms under communication constraints

Abstract Unmanned aerial vehicle (UAV) swarm systems are increasingly deployed in time-critical operations where multiple agents must coordinate autonomously under dynamic and often adversarial conditions. A persistent challenge in such deployments is the simultaneous presence of Byzantine faults, wherein compromised nodes inject malicious data into the swarm, and constrained communication bandwidth, which limits the fidelity of inter-agent information exchange. To address these dual challenges, a Zero-Trust Graph-Attentive Intelligence (ZT-GAT) framework is introduced. The proposed model employs an eight-head, 256-dimensional Graph Attention Network (GAT) backbone with a multi-signal zero-trust authentication layer fusing behavioral scoring, cryptographic attestation, and graph neighborhood analysis at a composite threshold of 0.75. Experiments were conducted on the UAVSwarm dataset augmented with a dedicated Byzantine and communication overlay spanning spoofing, replay, Sybil identity fabrication, and gradient poisoning attacks at node corruption levels of 5%, 10%, 20%, and 33%. In the model-level comparison, the proposed framework achieved a Byzantine F1-score of 96.8%, a trust classification accuracy of 97.9%, a consensus latency of 134.6 ms, and a packet delivery ratio (PDR) of 93.7%, with the evaluation covering swarm sizes reaching 50 agents, outperforming the strongest baseline by 6.1 percentage points in F1 at a computational cost of 1.91 GFLOPs per forward decision cycle. Separately, the unweighted mean Byzantine F1-score across the 12 attack–corruption conditions was 92.83%. These findings confirm that graph-attentive zero-trust integration provides a scalable and communication-aware foundation for Byzantine-resilient UAV swarm coordination.