LEO-Split: A Semi-Supervised Split Learning Framework over LEO Satellite Networks

TL;DR

LEO-Split tackles the combination of limited ground-station connectivity and labeling scarcity in LEO satellite networks by integrating semi-supervised learning with split learning. The framework introduces an auxiliary model to enable independent client-side updates during non-contact periods, an adaptive pseudo-labeling scheme to balance class and data quantity across satellites, and an adaptive activation interpolation strategy to mitigate server-side overfitting due to scarce activations. Empirical results on real Starlink traces and datasets GBSense and EuroSAT show faster convergence and higher accuracy than state-of-the-art SL baselines under both IID and non-IID settings. The approach enables scalable, robust DL in space environments and has potential for broader applications in multi-modal and large-language-model training over satellite networks.

Abstract

Recently, the increasing deployment of LEO satellite systems has enabled various space analytics (e.g., crop and climate monitoring), which heavily relies on the advancements in deep learning (DL). However, the intermittent connectivity between LEO satellites and ground station (GS) significantly hinders the timely transmission of raw data to GS for centralized learning, while the scaled-up DL models hamper distributed learning on resource-constrained LEO satellites. Though split learning (SL) can be a potential solution to these problems by partitioning a model and offloading primary training workload to GS, the labor-intensive labeling process remains an obstacle, with intermittent connectivity and data heterogeneity being other challenges. In this paper, we propose LEO-Split, a semi-supervised (SS) SL design tailored for satellite networks to combat these challenges. Leveraging SS learning to handle (labeled) data scarcity, we construct an auxiliary model to tackle the training failure of the satellite-GS non-contact time. Moreover, we propose a pseudo-labeling algorithm to rectify data imbalances across satellites. Lastly, an adaptive activation interpolation scheme is devised to prevent the overfitting of server-side sub-model training at GS. Extensive experiments with real-world LEO satellite traces (e.g., Starlink) demonstrate that our LEO-Split framework achieves superior performance compared to state-ofthe-art benchmarks.

Figures (15)

• Figure 1: Though SL may facilitate efficient satellite-GS collaborative model training, i) satellite-GS intermittent connectivity, ii) cumulative biases, and iii) data heterogeneity still pose significant challenges to real-world implementations.
• Figure 2: Limited contact time (b) and transmission rates (c) between LEO satellites and GS become a major bottleneck for training SL models (d).
• Figure 3: The SL performance against (a) different number of sub-models involved in model aggregation and (b) distinct labeling rates.
• Figure 4: The impact of data class imbalance (a) and quantity imbalance (b) on SL performance.
• Figure 5: An overview of LEO-Split architecture.