Partial Aggregation for Federated Learning with Differentiated Latency in LEO Satellite Networks

Federated learning (FL), which stands as a decentralized training paradigm, enables multiple users in different geographical locations to collaboratively learn a machine learning (ML) model without sharing their data, thus offering a solution to prevent privacy leakage in future communications networks. However, implementing FL in low earth orbit (LEO) satellite networks faces the challenge of latency optimization under the heterogeneity and energy consumption constraints of LEO satellites. To address this, we propose a partial centralized aggregation scheme to realize FL with differentiated latency by leveraging latency gaps to accelerate model convergence while introducing no extra latency in FL rounds. An optimization problem to maximize the number of users for partial centralized aggregation is formulated and solved efficiently within the energy consumption constraints of LEO satellites. Simulation results illustrate that our proposed scheme significantly expedites FL convergence compared to traditional FL methods.