SFL-LEO: Asynchronous Split-Federated Learning Design for LEO Satellite-Ground Network Framework

TL;DR

The paper addresses the challenge of training large models in highly dynamic LEO satellite-ground networks with limited on-satellite computation. It proposes SFL-LEO, a hybrid asynchronous split-federated learning framework that splits models between satellites and a ground station, uses an auxiliary network for local updates during satellite invisibility, and applies a layer-alignment strategy to accommodate heterogeneous satellite capabilities; aggregation at the ground station accounts for staleness via weighted updates. The approach integrates Split Learning and Federated Learning to maintain accuracy while drastically reducing satellite-to-ground data transfer, as demonstrated by Starlink-based experiments showing significant data savings with competitive accuracy compared to baseline SL/FL methods. The work includes identical and personalized split variants, a robust asynchronous SGD protocol, and a prototype implementation driven by real bandwidth traces, establishing SFL-LEO as a practical solution for efficient distributed learning in space-ground infrastructures. Overall, SFL-LEO offers a scalable means to leverage onboard compute and ground-station resources for timely, privacy-preserving model training in resource-constrained, dynamic satellite networks.

Abstract

Recently, the rapid development of LEO satellite networks spurs another widespread concern-data processing at satellites. However, achieving efficient computation at LEO satellites in highly dynamic satellite networks is challenging and remains an open problem when considering the constrained computation capability of LEO satellites. For the first time, we propose a novel distributed learning framework named SFL-LEO by combining Federated Learning (FL) with Split Learning (SL) to accommodate the high dynamics of LEO satellite networks and the constrained computation capability of LEO satellites by leveraging the periodical orbit traveling feature. The proposed scheme allows training locally by introducing an asynchronous training strategy, i.e., achieving local update when LEO satellites disconnect with the ground station, to provide much more training space and thus increase the training performance. Meanwhile, it aggregates client-side sub-models at the ground station and then distributes them to LEO satellites by borrowing the idea from the federated learning scheme. Experiment results driven by satellite-ground bandwidth measured in Starlink demonstrate that SFL-LEO provides a similar accuracy performance with the conventional SL scheme because it can perform local training even within the disconnection duration.

Figures (14)

• Figure 1: (a) A uniform constellation of $N=3$ satellites in $m=2$ orbits at two different altitudes.
• Figure 2: Visible patterns of satellites. Satellites 1 and 2 are at an altitude of $550$ Km, Satellites 3 and 4 are at an altitude of $340$ Km, and the GS is at the North Pole.
• Figure 3: Design Overview.
• Figure 4: The satellites have comparable computing capabilities and have the same split strategy.
• Figure 5: The timing diagram of the asynchronous SFL in an illustrative example with $N = 4, K = 1$.