Topology-Aware Two-Stage Federated Learning via Proxy Models for Sub-THz Heterogeneous LEO Communications

TL;DR

This paper introduces a topology-aware two-stage federated learning framework for LEO satellite networks, leveraging proxy models and dynamic aggregation to improve communication efficiency, resource utilization, and convergence.

Contribution

It proposes a novel two-stage FL framework with topology-aware aggregation and proxy models to address heterogeneity and staleness in LEO satellite communications.

Findings

Achieves 86.59%--90.57% test accuracy, outperforming baselines by 16.26%--19.80%.

Provides a 1.5x to 2.2x convergence speedup.

Extends contact windows using high-altitude platforms and Sub-THz links.

Abstract

Federated learning (FL) has emerged as a promising distributed training paradigm for Low Earth Orbit (LEO) networks by significantly reducing communication overhead. However, its deployment faces critical challenges, e.g., topology-induced model staleness, short contact windows, and unaddressed computing heterogeneity. To address these issues, a topology-aware two-stage FL framework is proposed in this paper. First, a multi-layer physical architecture utilizing high-altitude platforms (HAPs) and Sub-THz communications is designed to extend satellite-ground contact windows and enlarge available bandwidth. Second, a proxy-model-based approach is adopted to fully utilize heterogeneous resources and enable architecture-agnostic knowledge aggregation. Finally, building upon these foundations, a topology-aware two-stage aggregation mechanism is proposed as the central algorithmic design to overcome the topology-induced staleness. The mechanism dynamically partitions LEO satellites into localized groups based on their transient HAP coverage. Within each group, LEO satellites perform asynchronous aggregation at their associated HAP to naturally tolerate computational delays without penalizing faster nodes. Subsequently, a synchronous inter-group aggregation is executed among all HAPs at the Ground Station (GS) to strictly bound the maximum staleness and guarantee stable global convergence. Numerical results demonstrate the proposed framework extends contact windows and achieves 86.59%--90.57% test accuracy, outperforming the state-of-the-art heterogeneous baseline by 16.26\%--19.80\%. Furthermore, it achieves a 1.5x to 2.2x convergence speedup, which closely approaches the ideal upper bound.