Energy-Efficient UAV-Assisted Federated Learning in Wireless Networks.

With the proliferation of smart mobile devices and next-generation wireless communication technologies, federated learning (FL) has garnered significant attention as an emerging paradigm for privacy-preserving distributed model training. However, the traditional FL frameworks assume a static model aggregator such as the base station, which face multiple challenges including high energy consumption, frequent device dropout, and compromised model convergence. To address these issues, this study explores a novel FL framework called unmanned aerial vehicle (UAV)-assisted FL. The primary objective is to leverage UAVs as movable model aggregators, which collaborate with devices to minimize the energy consumption and ensure satisfactory convergence accuracy of FL. By adopting the distributed approximate newton (DANE) algorithm as the local optimizer, we first analyze the convergence of UAV-assisted FL and derive a device scheduling constraint to foster convergence. Subsequently, an optimization problem that aims at minimizing the total device energy consumption is formulated, which jointly optimizes the UAV trajectory, user selection, time slot length, and the uplink transmission power, CPU frequency, and local convergence accuracy of devices, while maintaining a desired global accuracy. This non-convex optimization problem is then decomposed into three subproblems and solved via the alternating direction method of multipliers (ADMM). Simulation results demonstrate that our proposed UAV-Assisted FL framework significantly reduces the total device energy consumption compared to baseline approaches and achieves a better balance with the model accuracy.