Dynamic Client Scheduling Enhanced Federated Learning for UAVs

Abstract

Although Federated Learning (FL) applied in Unmanned Aerial Vehicles (UAVs) offers substantial benefits, it also poses some challenges. These challenges arise primarily from the dynamic nature of UAV movements and the constraints imposed by limited wireless channel resources. This leads to the situation where only partial UAVs can participate in the FL process during each communication round, introducing the bias of the optimization objective that adversely impacts model accuracy. To address this issue, we introduce a Multi-action Q Network (MQN) for client scheduling, which selects suitable UAVs for each round, resolving the problems of the partial participation of UAVs. Furthermore, we propose a Gain-based Parameter Aggregation (GPA), which assigns a “contribution score" to each local model based on its contribution, correcting the bias of the optimization objective in FL. Simulation results demonstrate the effectiveness of the proposed methods.