Energy-Efficient Federated Learning Through UAV Edge Under Location Uncertainties

Federated Learning (FL) and Mobile Edge Computing (MEC) technologies alleviate the burden of deploying artificial intelligence (AI) on wireless devices with low computational capabilities. However, they also introduce energy consumption challenges in FL model training and data processing. In this paper, we employ Unmanned Aerial Vehicles (UAVs) to collect data from wireless devices and carry edge servers to assist the central server located at the base station in training FL model. We also consider the deviation of UAVs' locations to address its impact on network performance. Specifically, we formulate a robust joint optimization problem to minimize the energy consumption of UAVs, considering the computational resources, transmit power, transmission time, and FL model accuracy. Moreover, Gaussian-distributed uncertainties caused by deviation in UAV locations result in probabilistic constraints on data offloading. We initially employ the Bernstein-type inequality (BTI) to transform probabilistic constraints into deterministic forms. Subsequently, we adopt the Block Coordinate Descent (BCD) to separate the problem into three subproblems. Simulation results demonstrate a significant reduction in energy consumption and superiority in robustness.