SFedXSL: Semi-Synchronous Federated Cross-Sharpness Learning for UAV Swarm

Federated learning (FL) emerges as an innovative approach to manage a collection of client UAVs in order to co-train machine-learning models that are readily integrated into an Unmanned Aerial Vehicle (UAV) swarm. However, due to more erratic communication conditions than in terrestrial wireless networks, synchronous aggregation—which is utilized in traditional FL—is no longer feasible in UAV swarm. Additionally, because of the various deployment zones or specifications of UAVs, the data gathered by them are frequently heterogeneous. A significant amount of unlabeled data will be collected by the UAV swarm in light of its new flight trajectory and unseen scenarios. To address these issues, we propose a unique Semi-synchronous Federated Cross-Sharpness Learning (SFedXSL) framework to tackle the problems of semisynchronous devices and unlabeled data. This framework incorporates unsupervised pre-training and client UAV clustering scheduling. These proposed techniques aim to unlock the full potential of unlabeled and labeled user data, expediting the training process. Simulation results demonstrate that our proposed algorithm surpasses state-of-the-art FL techniques in terms of objection recognition accuracy and service latency.