Scheduling for Ground-Assisted Federated Learning in LEO Satellite Constellations

Abstract

Distributed training of machine learning models directly on satellites in low Earth orbit (LEO) is considered. Based on a federated learning (FL) algorithm specifically targeted at the unique challenges of the satellite scenario, we design a scheduler that exploits the predictability of visiting times between ground stations (GS) and satellites to reduce model staleness. Numerical experiments show that this can improve the convergence speed by a factor three.

AsyncFLEO: Asynchronous Federated Learning for LEO Satellite Constellations with High-Altitude Platforms

“…FedSat adapts FedAvg [4] to an asynchronous setting, where it averages the received satellite models based on their visibility order, assuming regular satellite visits to the GS. In response to this ideal consideration, they proposed another work [14], FedSatSchedule, that considers the location of the GS can be anywhere.…”

“…• Synchronous FL approaches: FedAvg [4], FedHAP [8], FedISL [7], and DSFL [11]. • Asynchronous FL approaches: FedSatSchedule [14],…”

“…To evaluate the performance of NomaFedHAP against baseline approaches, we focus on image classification. We employ commonly used model training datasets including MNIST, CIFAR-10, and CIFAR-100, which are frequently utilized in various FL-SatCom studies [6], [8], [14], [17]. In addition, despite the lack of space application datasets, we utilize a real dataset of high-resolution satellite images called DeepGlobe for road extraction to provide a realistic evaluation of NomaFedHAP as well as demonstrate its applicability to real-world scenarios.…”

“…However, their usability in realistic scenarios is limited due to their oversimplified assumption that satellites should visit the GS in a regular manner or overlook the sink satellite's visibility period. When the authors of FedSat omitted this assumption by developing FedSatSchedule [14], the convergence speed is doubled to 48 hours, and the accuracy is reduced by 14-24% in comparison with FedSat on different datasets. Although FedSatSchedule offers more realistic consideration, traditional FedAsync [5] can achieve similar accuracy for the same convergence period.…”