Importance-Aware Client Scheduling and Resource Allocation for Federated Learning in UAV Networks

Adopting federated learning (FL) in unmanned aerial vehicle (UAV) networks is a promising paradigm, which can empower UAV networks with enhanced intelligence to support complex applications. Considering the imbalanced data properties, limited energy and unstable wireless connection of UAVs, an effective client scheduling scheme is critical for the design of efficient FL. In this paper, in order to properly consider the priority criteria, we first propose two importance metrics from the perspectives of data attributes and local updates, namely data importance measurement (DIM) and gradient importance measurement (GIM). Then, take into account DIM and GIM, an optimization problem is formulated to jointly optimize the client scheduling, computation and communication of UAVs. Due to the non-convex nature of this problem, we decompose it into two sub-problems and derive their optimal closed-form solutions. Simulations demonstrate that, compared to benchmark schemes, our proposal ensures better performance on test accuracy, convergence and energy saving.