Heterogeneity‐Aware Asynchronous Federated Learning for UAV Cooperative Perception

ABSTRACT Unmanned aerial vehicle ( UAV ) swarm cooperative perception faces critical challenges including system heterogeneity, non‐independent and identically distributed (non‐IID) data distributions and limited communication resources, which render conventional synchronous federated learning (FL) impractical. We propose a heterogeneity‐aware asynchronous federated learning (HAFL) framework that adaptively adjusts each UAV client's local training workload through an elastic aggregation window, enabling more nodes to participate in global aggregation without excessive model staleness. A hierarchical aggregation mechanism groups UAV nodes by data distribution similarity and applies gradient‐directed pruning and amplification to accelerate convergence under non‐IID conditions. A dynamic edge‐weight penalty jointly considers model staleness, local accuracy and participation frequency. Compared with FedAvg and FedBuff baselines, HAFL improves classification accuracy by up to 3.27%, reduces system cost by up to 23.2% and increases average per‐client participation rate by 0.21. Experiments demonstrate that HAFL significantly outperforms existing asynchronous FL baselines in client participation rate, global accuracy, convergence speed, system cost and energy efficiency.