[Paper Review] FedSN: A Federated Learning Framework over Heterogeneous LEO Satellite Networks

Lin Zheng, Zhe Chen+4 more|arXiv (Cornell University)|Nov 2, 2023

Satellite Communication Systems4 citations

TL;DR

FedSN is a novel federated learning framework for heterogeneous Low Earth Orbit (LEO) satellite networks that addresses computing, memory, and communication heterogeneity, limited uplink bandwidth, and model staleness through a sub-structure scheme and pseudo-synchronous model aggregation. It achieves higher accuracy and lower communication and computation overhead than state-of-the-art methods in remote sensing and space modulation tasks.

ABSTRACT

Recently, a large number of Low Earth Orbit (LEO) satellites have been launched and deployed successfully in space by commercial companies, such as SpaceX. Due to multimodal sensors equipped by the LEO satellites, they serve not only for communication but also for various machine learning applications, such as space modulation recognition, remote sensing image classification, etc. However, the ground station (GS) may be incapable of downloading such a large volume of raw sensing data for centralized model training due to the limited contact time with LEO satellites (e.g. 5 minutes). Therefore, federated learning (FL) has emerged as the promising solution to address this problem via on-device training. Unfortunately, to enable FL on LEO satellites, we still face three critical challenges that are i) heterogeneous computing and memory capabilities, ii) limited uplink rate, and iii) model staleness. To this end, we propose FedSN as a general FL framework to tackle the above challenges, and fully explore data diversity on LEO satellites. Specifically, we first present a novel sub-structure scheme to enable heterogeneous local model training considering different computing, memory, and communication constraints on LEO satellites. Additionally, we propose a pseudo-synchronous model aggregation strategy to dynamically schedule model aggregation for compensating model staleness. To further demonstrate the effectiveness of the FedSN, we evaluate it using space modulation recognition and remote sensing image classification tasks by leveraging the data from real-world satellite networks. Extensive experimental results demonstrate that FedSN framework achieves higher accuracy, lower computing, and communication overhead than the state-of-the-art benchmarks and the effectiveness of each components in FedSN.

Motivation & Objective

• Address the challenge of training effective machine learning models on LEO satellites due to limited ground station contact time and high data volume.
• Overcome resource heterogeneity among LEO satellites, including varying computing, memory, and communication capabilities.
• Mitigate model staleness caused by intermittent connectivity and asynchronous updates in satellite FL.
• Reduce communication overhead by optimizing model aggregation under low uplink bandwidth constraints.
• Enable efficient, privacy-preserving, on-device training across diverse satellite constellations without relying on inter-satellite links.

Proposed method

• Proposes a sub-structure scheme that customizes local model architectures based on individual satellite’s computing, memory, and communication constraints.
• Introduces a sub-structure distribution mechanism that balances training load across satellites with different capabilities.
• Employs a sub-structure aggregation strategy to reduce intra-group model staleness by prioritizing updates from less stale models.
• Designs a pseudo-synchronous model aggregation strategy that uses a discrepancy-aware weighting function based on model parameter differences to dynamically weight incoming updates.
• Utilizes a buffer-based aggregation mechanism to store and schedule model updates, enabling asynchronous yet coordinated aggregation to reduce inter-group staleness.
• Incorporates time- and discrepancy-dependent weighting to balance the influence of stale and recent models, improving convergence and generalization.

Experimental results

Research questions

• RQ1 How can federated learning be effectively scaled across LEO satellites with heterogeneous computing, memory, and communication capabilities?
• RQ2 To what extent can model staleness be reduced in LEO satellite networks when contact with ground stations is intermittent and short-lived?
• RQ3 Can a discrepancy-aware model aggregation strategy outperform time-only-based staleness compensation in satellite FL?
• RQ4 How does sub-structure customization improve training efficiency and model accuracy under resource constraints?
• RQ5 What is the impact of pseudo-synchronous aggregation on convergence speed and test accuracy in real-world satellite datasets?

Key findings

• FedSN achieves higher test accuracy than state-of-the-art benchmarks on both space modulation recognition and remote sensing image classification tasks.
• The sub-structure customization and distribution mechanism significantly reduces training imbalance and improves model convergence across heterogeneous satellites.
• The pseudo-synchronous model aggregation strategy accelerates convergence and improves accuracy by 3.2% on average compared to FedAsync and 2.1% over SS+FedAsync on the GBsense dataset.
• Ablation studies confirm that each component—sub-structure customization, distribution, aggregation, and pseudo-synchronous aggregation—contributes meaningfully to performance gains.
• The framework reduces communication and computation overhead by optimizing model size and update frequency based on satellite capabilities.
• The discrepancy-based weighting function in model aggregation effectively mitigates inter-group staleness, especially when satellites have long orbital periods and irregular contact windows.