Graph Spectral Client Scheduling for Reliable Federated Learning in Safety

Abstract
Federated learning (FL) over low Earth orbit (LEO) satellite constellations faces a fundamental safety challenge. Intermittent satellite visibility causes standard client scheduling methods to systematically exclude certain geographic regions, producing ground stations with zero recall that render entire populations invisible to safety critical event detection. This work demonstrates that existing schedulers, including random selection, loss based prioritization, and operational heuristics such as maximum visibility and geographic diversity, routinely produce this catastrophic failure mode under realistic orbital dynamics, yet the failure remains hidden when only aggregate metrics are reported. Standard evaluation, relying solely on aggregate metrics such as average test accuracy and convergence speed, conceals catastrophic localized failures that disaggregated per node safety metrics reveal. To substantially reduce this failure, a graph spectral scheduling framework called Graph Centric Periodic Scheduling (GCPS) is introduced. GCPS preserves the algebraic connectivity of the training subgraph by jointly optimizing betweenness centrality, participation diversity, eigenvector contribution to the Fiedler vector, and node degree, with weights dynamically adapted through a validation only multi armed bandit exploring seven Pareto optimal weight configurations. The central finding establishes that GCPS substantially reduces zero recall nodes to 0 out of 5 trials (0%), compared to 5 out of 5 trials (100%) for geographic diversity scheduling and 2 out of 5 trials (40%) for state of the art fairness methods, namely Agnostic Federated Learning (AFL) and q Fair Federated Learning (q-FFL). This substantial improvement is unmatched by any of the nine baseline methods, while GCPS also significantly reduces false negative rates (\(p < 0.01\) versus geographic diversity, graph weighted sampling, q-FFL, AFL, and Oort), maintains competitive standard accuracy, and demonstrates robustness to topology mismatch, differential privacy noise, and label flipping attacks. Beyond the specific satellite domain, this work argues that minimum per node performance guarantees must become a primary evaluation criterion for safety critical FL, as aggregate metrics conceal catastrophic localized failures that graph spectral scheduling is uniquely positioned to prevent. It is further demonstrated that commonly used fairness indices, with Jain’s index exceeding 0.98 for all methods, measure participation equity rather than performance equity, and therefore fail to detect zero recall failures, a critical distinction for safety critical deployment.
Similar content being viewed by others
Data Availability
No datasets were generated or analysed during the current study.
References
Abtahi, N., Chaman, U. M., & Kabir, M. S. (2026). Federated learning in healthcare: A comprehensive survey on privacy, scalability and clinical applications. ICT Express. https://doi.org/10.1016/j.icte.2026.05.011
Choi, Chang-Sik & Ku, Bon-Jun & Baccelli, Francois. (2024). Stochastic geometry and dynamical system analysis of walker constellation networks. https://doi.org/10.48550/arXiv.2412.01610
de Curtò, J., de Zarzà, I., Roig, G., Cano, J.-C., Manzoni, P., & Calafate, C. (2023). LLM-Informed Multi-Armed Bandit Strategies for Non-Stationary Environments. Electronics., 12, 2814. https://doi.org/10.3390/electronics12132814
Dobariya, R., & Hobiger, T. (2026). Automatic scheduling of satellite tracking tasks by means of policy-gradient reinforcement learning and transformer-based pointer networks. CEAS Space Journal. https://doi.org/10.1007/s12567-026-00726-y
Fu, L., Zhang, H., Gao, G., Zhang, M., & Liu, X. (2023). Client selection in federated learning: Principles, challenges, and opportunities. IEEE Internet of Things Journal. https://doi.org/10.1109/JIOT.2023.3299573
Ji, B., Han, J., Kim, D., Yun, J., & Joo, C. (2026). Inter-satellite link technologies and applications in Low Earth Orbit satellite networks. ICT Express, 12(3), 576–592. https://doi.org/10.1016/j.icte.2026.03.001
Ji, X., Zhang, Y., Zheng, K., et al. (2025). Federated asynchronous graph attention network with structural semantic embedding for multi-label graph classification. Scientific Reports, 15, Article 38618. https://doi.org/10.1038/s41598-025-22500-6
Johnson, Dr. (2024). Wireless sensor networks for earthquake and natural disaster detection. American Journal of Sensor Networks and Wireless Communications, 5, 1–12. https://doi.org/10.71465/ajsnwc.3193
Kim, D., Woo, H., & Lee, Y. (2024). Addressing bias and fairness using fair federated learning: A systematic literature review. https://doi.org/10.20944/preprints202410.1060.v1
Lagunas, E., & Ottersten, B. (2024). Low-earth orbit satellite constellations for global communication network connectivity. Nature Reviews Electrical Engineering. https://doi.org/10.1038/s44287-024-00088-9
Li, Y., Liu, L., Li, H., Liu, W., Chen, Y., Zhao, W., Wu, J., Wu, Q., Liu, J., and Lai, Z. (2024). Stable Hierarchical Routing for Operational LEO Networks. In Proceedings of ACM Conference, pp. 296–311. https://doi.org/10.1145/3636534.3649362
Li, Y., Zhang, C., Jia, C., Li, X., & Zhu, Y. (2019). Joint optimization of workforce scheduling and routing for restoring a disrupted critical infrastructure. Reliability Engineering & System Safety., 191, Article 106551. https://doi.org/10.1016/j.ress.2019.106551
Mateo-Garcia, G., Veitch-Michaelis, J., Purcell, C., Longépé, N., Reid, S., Anlind, A., Bruhn, F., Parr, J., & Mathieu, P. (2023). In-orbit demonstration of a re-trainable machine learning payload for processing optical imagery. Scientific Reports. https://doi.org/10.1038/s41598-023-34436-w
Paterson, C., Hawkins, R., Picardi, C., Jia, Y., Calinescu, R., & Habli, I. (2025). Safety assurance of Machine Learning for autonomous systems. Reliability Engineering & System Safety., 264, Article 111311. https://doi.org/10.1016/j.ress.2025.111311
Pavel, M. I., Hu, S., Pratama, M., & Kowalczyk, R. (2026). Onboard Optimization and Learning: A Survey. IEEE Access, 14, 33072–33104. https://doi.org/10.1109/ACCESS.2026.3664956
Pelton, J.N., Laufer, R. (2020). Commercial Small Satellites for Business Constellations Including Microsatellites and Minisatellites. In: Pelton, J.N., Madry, S. (eds) Handbook of Small Satellites. Springer, Cham. https://doi.org/10.1007/978-3-030-36308-6
Satish, S., Gonaygunta, H., Yadulla, A. R., Kumar, D., Maturi, M. H., Meduri, K., De La Cruz, E., Nadella, G. S., & Sajja, G. S. (2025). Forecasting the unseen: Enhancing tsunami occurrence predictions with machine-learning-driven analytics. Computers, 14(5), Article 175. https://doi.org/10.3390/computers14050175
Shinde, S. S., Naseh, D., Tarchi, D., & Fischione, C. (2026). LEO satellites accelerating edge intelligence: Model transfer and similarity-aware initialization for federated learning. IEEE Internet of Things Magazine. https://doi.org/10.1109/MIOT.2026.3700215
Wang, B., Li, A., Li, H. H., & Chen, Y. (2020). GraphFL: A Federated Learning Framework for Semi-Supervised Node Classification on Graphs. IEEE International Conference on Data Mining (ICDM), 2022, 498–507. https://doi.org/10.1109/ICDM54844.2022.00060
Waref, D., Alayary, Y., & Abd El Ghany, M. A. (2026). Review on federated optimization in multi-tier architectures. Journal of King Saud University - Computer and Information Sciences, 38, Article 101. https://doi.org/10.1007/s44443-026-00510-2
Zaidyn, M., Akhtanov, S., Turlykozhayeva, D., Temesheva, S., Akhmetali, A., Skabylov, A., & Ussipov, N. (2025). The role of fractal dimension in wireless mesh network performance. Scientific Reports, 15(1), 42149. https://doi.org/10.48550/arXiv.2506.19366
Author information
Authors and Affiliations
Contributions
I am the only author.
Corresponding author
Additional information
Editors: Bruno Casella, Linara Adilova, Michael Kamp.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Ahmad, B. Graph Spectral Client Scheduling for Reliable Federated Learning in Safety-Critical LEO Satellite Networks. Mach Learn 115, 214 (2026). https://doi.org/10.1007/s10994-026-07159-y
Received:
Revised:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1007/s10994-026-07159-y

Facts Only

* Federated learning over LEO satellite constellations faces a safety challenge due to intermittent visibility causing exclusion of geographic regions from scheduling.
* Standard schedulers (random selection, loss-based prioritization, maximum visibility heuristics) routinely produce zero recall ground stations under realistic orbital dynamics.
* Aggregate metrics like average test accuracy conceal localized failures.
* Graph Centric Periodic Scheduling (GCPS) is introduced as a new scheduling framework.
* GCPS optimizes joint metrics: betweenness centrality, participation diversity, eigenvector contribution to the Fiedler vector, and node degree.
* Optimization weights in GCPS are dynamically adapted via a multi-armed bandit exploring seven Pareto optimal configurations.
* GCPS reduced zero recall nodes to 0 out of 5 trials (0%) compared to 5 out of 5 trials (100%) for geographic diversity scheduling.
* GCPS reduced false negative rates below 0.01 compared to methods like geographic diversity, graph weighted sampling, q-FFL, AFL, and Oort.
* GCPS maintains competitive standard accuracy.
* Fairness indices, such as Jain’s index exceeding 0.98 for all methods, measure participation equity rather than performance equity and fail to detect zero recall failures.

Executive Summary

Existing scheduling methods in federated learning over LEO satellite constellations often exclude specific geographic regions due to intermittent satellite visibility, leading to ground stations with zero recall for safety-critical events and rendering populations invisible to detection. Standard evaluation relying only on aggregate metrics like average test accuracy or convergence speed fails to reveal these localized failures. The research introduces Graph Centric Periodic Scheduling (GCPS), a novel framework designed to address this by optimizing scheduling based on graph spectral properties, including betweenness centrality, participation diversity, eigenvector contribution to the Fiedler vector, and node degree. GCPS employs a multi-armed bandit approach to dynamically select optimal weighting configurations for these factors. Empirical results show that GCPS substantially reduces zero recall nodes across trials compared to baseline methods, achieving 0% in five out of five trials for geographic diversity scheduling, and significantly reducing false negative rates. The findings suggest that minimum per-node performance guarantees are necessary for safety-critical federated learning systems, as aggregate metrics mask localized catastrophic failures.

Full Take

The central tension in this work lies between the utility of aggregate performance metrics and the necessity of localized safety guarantees in complex systems. The finding that commonly used fairness indices measure superficial participation equity rather than genuine performance equity highlights a critical gap: operational metrics prioritize smooth, high-level convergence over the existential risk posed by isolated failures—specifically, zero recall nodes. This points toward a systemic failure in the validation paradigm for safety-critical AI deployments where localized catastrophic events are deliberately obscured by aggregation.
The introduction of GCPS moves beyond traditional heuristic scheduling to leverage graph spectral analysis. By optimizing properties like betweenness centrality and Fiedler vector contribution, the framework attempts to preserve the underlying algebraic connectivity of the training subgraph, suggesting that systemic resilience should be derived from the network topology itself rather than superficial performance averages. The superiority demonstrated by GCPS over methods like Geographic Diversity and state-of-the-art fairness methods (AFL, q-FFL) implies that structural awareness—understanding how nodes connect within the operational graph—is a prerequisite for safety assurance in distributed systems.
The broader argument extends to the philosophy of evaluation: safety demands disaggregated metrics. If an agent is designed for safety, its performance must be assessed at the most granular level possible, not just through smoothed aggregates. The failure of Jain’s index to capture this distinction suggests that established fairness tools are insufficient when stakes involve physical safety and distributed infrastructure integrity. The work posits a necessary shift: minimum per-node performance guarantees should supersede aggregate metrics as the primary criterion for deployment readiness in safety-critical federated learning. What structure exists outside the mathematical optimization of these spectral properties that must be integrated into safety-first engineering principles?

Sentinel — Human

Confidence

This analysis details a novel mathematical scheduling framework designed to prevent catastrophic failure in Federated Learning systems over LEO satellites by focusing on local safety guarantees rather than aggregate metrics.

Signals Detected

Sentence length variance shows moderate variation; complex topic management suggests human authoring.

The argument flows logically from a specific technical problem (scheduling failure) to a novel solution (GCPS) and broad implications (safety guarantees), indicating focused expertise.

Citations are specific, linking to highly relevant, recent academic literature in LEO, FL, and graph theory, suggesting deep domain knowledge rather than simple LLM aggregation.

The mathematical results cited (e.g., 'reduces zero recall nodes to 0 out of 5 trials (0%)') are highly specific, typical of a novel empirical study, though this is not proof against synthetic generation.

Human Indicators

The dense layering of specialized concepts (Graph Spectral Scheduling, Graph Centric Periodic Scheduling, multi-armed bandit exploration) integrated into a specific aerospace/ML context points toward an author operating within the immediate expert domain.

The distinction drawn between participation equity and performance equity regarding fairness indices suggests a nuanced philosophical stance beyond pure data recitation.