SafeLearning: mobility-aware client selection in federated edge learning over UAV networks

One promising way to enable distributed intelligence at the edge while protecting data privacy is to integrate Federated Learning (FL) with Unmanned Aerial Vehicles (UAV) networks. Using FL enables each UAV to cooperatively train a global model without sharing raw data, especially in UAV swarms used for surveillance, monitoring, or emergency response missions. But choosing the best clients (UAVs) for every training cycle is made extremely difficult by the dynamic and diverse character of UAV environments. These difficulties are brought on by things like fluctuating connectivity, shifting patterns of movement, and energy limitations. In this work, we investigate the problem of client selection for Federated Edge Learning in UAV networks. We first present a taxonomy of existing selection strategies, considering criteria such as model performance and UAV mobility. Then, we propose SafeLearning, an adaptive client selection framework that integrates both mobility-awareness and distance with speed to enhance learning efficiency and model accuracy. Simulations demonstrate that our method significantly improves convergence speed and reduces client dropout, while maintaining high model performance in dynamic UAV scenarios.