Overhead Reduction in UAV-Assisted Federated Learning with Fast-Varying Environment

Thanks to its communication efficiency, low latency, and privacy-preserving nature, federated learning (FL) has emerged as a promising learning paradigm in the unmanned aerial vehicle (UAV)-enabled networks, where the UAVs support various applications. Despite its great potential, FL over a wireless network suffers from excessive communication overhead, particularly when the data distribution changes frequently. This drawback, which results from the necessity of frequent model updates, degrades the network performance severely. To address this challenge, we propose two methods based on offline and online Bayesian changepoint detection. In such approaches, the FL mechanism is triggered only upon observing abrupt changes in data distribution at a sufficiently large number of clients that can severely degrade the model accuracy. Our results show that both approaches yield near-optimal FL performance while significantly reducing the overhead.