Joint scheduling and resource allocation for federated learning in SWIPT-enabled micro UAV swarm networks

Abstract

Micro-UAV swarms usually generate massive data when performing tasks. These data can be harnessed with various machine learning (ML) algorithms to improve the swarm's intelligence. To achieve this goal while protecting swarm data privacy, federated learning (FL) has been proposed as a promising enabling technology. During the model training process of FL, the UAV may face an energy scarcity issue due to the limited battery capacity. Fortunately, this issue is potential to be tackled via simultaneous wireless information and power transfer (SWIPT). However, the integration of SWIPT and FL brings new challenges to the system design that have yet to be addressed, which motivates our work. Specifically, in this paper, we consider a micro-UAV swarm network consisting of one base station (BS) and multiple UAVs, where the BS uses FL to train an ML model over the data collected by the swarm. During training, the BS broadcasts the model and energy simultaneously to the UAVs via SWIPT, and each UAV relies on its harvested and battery-stored energy to train the received model and then upload it to the BS for model aggregation. To improve the learning performance, we formulate a problem of maximizing the percentage of scheduled UAVs by jointly optimizing UAV scheduling and wireless resource allocation. The problem is a challenging mixed integer nonlinear programming problem and is NP-hard in general. By exploiting its special structure property, we develop two algorithms to achieve the optimal and suboptimal solutions, respectively. Numerical results show that the suboptimal algorithm achieves a near-optimal performance under various network setups, and significantly outperforms the existing representative baselines. considered.