Staleness-Aware UAV-Assisted Over-the-Air Asynchronous Federated Learning

Abstract

The use of unmanned aerial vehicles (UAVs) to facilitate federated learning (FL) has attracted growing attention, as UAVs can mitigate key limitations of terrestrial FL systems by enabling reliable participation of devices located at the network edge. This capability is particularly important for FL systems employing over-the-air (OTA) aggregation, where devices experiencing weak wireless channels can severely distort the aggregated updates. However, the limited coverage of UAVs introduces staleness issues, since devices may not be able to participate synchronously, leading to degraded learning performance. To address these challenges, we propose a staleness-aware UAV-assisted OTA asynchronous FL framework that jointly optimizes transceiver design, device scheduling, and UAV trajectory. We first establish a convergence analysis that explicitly characterizes the impact of staleness on learning performance. Building on this analysis, we formulate a joint optimization problem that minimizes the performance loss due to staleness and OTA communication errors. We propose an alternating optimization algorithm to solve this problem, and analyze its convergence and computational complexity. To cope with time-varying wireless fading in practical UAV operations, we further propose an online transceiver adjustment mechanism that dynamically refines the transceiver parameters during FL execution. Extensive simulation results based on benchmark learning datasets demonstrate that the proposed framework achieves faster convergence and higher accuracy than existing schemes.