CoDisc: Collaborative Discrepancy-Aware Federated Learning for UAV Swarms

Federated learning (FL) provides a privacy-preserving paradigm for distributed intelligence, making it particularly suitable for unmanned aerial vehicle (UAV) swarms that collaboratively collect data and perform sensing tasks. However, the heterogeneity in sensing capabilities and local data distributions across UAVs often results in imbalanced and divergent contributions, degrading aggregation efficacy and limiting the generalization ability of the global model. To address this challenge, we propose CoDisc, a Collaborative Discrepancy level-aware FL framework for UAV swarms. CoDisc leverages both data distribution discrepancies and the feature-space complementarity among UAVs. It systematically quantifies the informativeness of each UAV’s update and explicitly distinguishes between harmful distribution divergence and beneficial feature-space complementarity. By integrating these aspects into a collaborative discrepancy level, CoDisc dynamically assigns adaptive aggregation weights guided by the assessed informativeness of each update. We theoretically characterize CoDisc by deriving a bound on the global objective and an analytic expression for aggregation weights that jointly account for dataset size and collaborative discrepancy among UAVs. Extensive evaluations on five real-world datasets demonstrate that CoDisc consistently outperforms nine state-of-the-art baselines. It achieves average improvements of 1.37% (NIID-1) and 1.15% (NIID-2) on challenging datasets such as HAM10000, CIFAR-10 & CIFAR-100, and CINIC-10, while maintaining steady gains of 0.6% on simpler datasets like Fashion-MNIST and AGNEWS. These results validate the effectiveness and broad applicability of CoDisc’s collaborative discrepancy-aware aggregation in FL for UAV swarms.