Sparse Incremental Aggregation in Satellite Federated Learning

TL;DR

This paper explores a bandwidth-efficient federated learning approach for satellite constellations using incremental aggregation and gradient sparsification, significantly improving communication efficiency in low Earth orbit networks.

Contribution

It introduces the application of gradient sparsification to satellite federated learning with incremental aggregation, enhancing bandwidth efficiency in LEO satellite networks.

Findings

Over 4x increase in bandwidth efficiency with more satellites

Gradient sparsification effectively reduces communication load

In-network incremental aggregation improves FL scalability

Abstract

This paper studies Federated Learning (FL) in low Earth orbit (LEO) satellite constellations, where satellites are connected via intra-orbit inter-satellite links (ISLs) to their neighboring satellites. During the FL training process, satellites in each orbit forward gradients from nearby satellites, which are eventually transferred to the parameter server (PS). To enhance the efficiency of the FL training process, satellites apply in-network aggregation, referred to as incremental aggregation. In this work, the gradient sparsification methods from [1] are applied to satellite scenarios to improve bandwidth efficiency during incremental aggregation. The numerical results highlight an increase of over 4 x in bandwidth efficiency as the number of satellites in the orbital plane increases.