pith. sign in

arxiv: 2604.15779 · v1 · submitted 2026-04-17 · 💻 cs.DC

CroSatFL: Energy-Efficient Federated Learning with Cross-Aggregation for Satellite Edge Computing

Nan Yang , Bahman Javadi , Rodrigo Neves Calheiros , David Boland , Philip Leong This is my paper

Pith reviewed 2026-05-10 07:58 UTC · model grok-4.3

classification 💻 cs.DC
keywords federated learningsatellite edge computingenergy efficiencyLEO constellationson-orbit aggregationinter-satellite linksstraggler mitigation
0
0 comments X

The pith

CroSatFL runs federated learning on LEO satellites with only two ground station contacts by shifting all iterative training and aggregation to on-orbit laser links.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces CroSatFL to make federated learning practical in energy-limited satellite edge environments. It keeps the ground station out of every training round by handling local updates and intermediate aggregations entirely among satellites. Three mechanisms support this: StarMask builds clusters that match data sizes to each satellite's compute power, Skip-One drops at most one slow client per round to cut energy without harming fairness, and random-k cross-aggregation mixes models across clusters at negligible extra cost. Simulations on a Walker-Delta constellation show the design slashes ground-station communications by more than two orders of magnitude and transmission energy by roughly 6 times while matching baseline accuracy and converging faster.

Core claim

By confining ground-station exchanges to initialization and final collection and performing all iterative steps on orbit, CroSatFL reduces GS communication count by over two orders of magnitude and GS transmission energy by about 6x relative to GS-centric and on-orbit baselines, while achieving competitive accuracy and faster convergence under realistic LEO dynamics.

What carries the argument

CroSatFL's hierarchical on-orbit aggregation that uses StarMask to form LISL-feasible clusters aligned with heterogeneous hardware, Skip-One to skip transient stragglers, and random-k cross-aggregation for lightweight topology-aware mixing.

If this is right

  • Ground station bandwidth and energy budgets become far less limiting factors for satellite federated learning.
  • Training rounds can complete faster because iterative exchanges stay on high-bandwidth laser links instead of waiting for ground round-trips.
  • Heterogeneous satellite hardware can still participate without one slow node dominating every round's energy cost.
  • Overall constellation energy budgets support longer or more frequent learning tasks without added solar-panel or battery capacity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same two-contact pattern could extend to other distributed training workloads that tolerate delayed global synchronization.
  • If laser link reliability holds, the framework suggests on-orbit computing clusters could handle many latency-sensitive edge tasks beyond machine learning.
  • Future hardware designs for satellites might prioritize low-power inter-satellite transceivers over repeated ground-link hardware.

Load-bearing premise

The end-to-end energy model and Walker-Delta constellation simulation accurately capture real LEO dynamics, heterogeneous hardware power draw, and laser link feasibility without unmodeled interference or hardware failures.

What would settle it

Running CroSatFL on actual LEO hardware and comparing measured energy use and communication counts against the simulation predictions under varying orbital conditions.

Figures

Figures reproduced from arXiv: 2604.15779 by Bahman Javadi, David Boland, Nan Yang, Philip Leong, Rodrigo Neves Calheiros.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework for CroSatFL. StarMask first forms constraint-aware clusters over heterogeneous satellites. Skip-One then [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Convergence comparison on MNIST, CIFAR-10, and EuroSAT using ResNet-18 under IID settings. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Convergence comparison on MNIST, CIFAR-10, and EuroSAT using ResNet-18 under non-IID settings. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of total energy consumption and end-to-end training time for a common target accuracy. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Single edge round computation energy and time comparison under [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
read the original abstract

Low Earth Orbit (LEO) mega-constellations extend the cloud-to-edge continuum into space, enabling satellite edge computing. However, Federated Learning (FL) in this environment is fundamentally energy-constrained due to dynamic inter-satellite connectivity, heterogeneous onboard computing hardware, and strict power budgets. We propose CroSatFL, a sustainable on-orbit hierarchical FL framework that reduces end-to-end energy across computation and communication while maintaining strong training performance under realistic LEO dynamics. CroSatFL keeps the ground station (GS) off the iterative loop by performing all local training and intermediate aggregations on orbit, requiring only two GS communication phases: one for initialization and one for final model collection. This sharply reduces repeated use of bandwidth-limited and energy-expensive GS links and shifts iterative exchanges to laser inter-satellite links (LISLs). CroSatFL integrates three energy-aware mechanisms: StarMask forms LISL-feasible clusters that align data volume with heterogeneous CPU/GPU capability, Skip-One mitigates transient stragglers by skipping at most one slow client per cluster to lower round energy and latency while preserving long-term fairness, and random-k cross-aggregation enables lightweight topology-aware cross-cluster mixing without extending round duration. Using an end-to-end energy model with a realistic Walker-Delta constellation, we show that CroSatFL reduces GS communication count by over two orders of magnitude and GS transmission energy by about 6x relative to GS-centric and on-orbit baselines, while achieving competitive accuracy and faster convergence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript presents CroSatFL, a hierarchical federated learning framework for LEO satellite edge computing that performs all local training and intermediate aggregations on-orbit using laser inter-satellite links (LISLs), limiting ground station (GS) involvement to initialization and final model collection phases only. It introduces three mechanisms—StarMask for forming LISL-feasible clusters aligned with data volume and heterogeneous CPU/GPU capabilities, Skip-One for mitigating transient stragglers by skipping at most one slow client per cluster, and random-k cross-aggregation for lightweight topology-aware mixing across clusters—evaluated via an end-to-end energy model on a realistic Walker-Delta constellation. The work claims over two orders of magnitude reduction in GS communication count and approximately 6x reduction in GS transmission energy relative to GS-centric and on-orbit baselines, while achieving competitive accuracy and faster convergence.

Significance. If the simulation results and energy model hold, this could represent a meaningful advance in sustainable on-orbit machine learning by addressing energy and connectivity constraints in mega-constellations, potentially enabling more efficient FL deployments in space edge computing scenarios. The hierarchical on-orbit design and specific mechanisms for clustering, straggler handling, and cross-aggregation provide concrete contributions to the intersection of distributed systems and satellite networks.

major comments (3)
  1. [Evaluation / Energy Model] The central quantitative claims (over 100x GS communication reduction and 6x energy savings) rest entirely on the fidelity of the end-to-end energy model and Walker-Delta simulation described in the evaluation. The manuscript must provide the explicit equations and parameter values used for onboard computation energy (CPU/GPU power draw) and LISL communication energy, including all assumptions on link feasibility, interference, and hardware heterogeneity, as these directly determine the reported gains relative to baselines.
  2. [Proposed Mechanisms / Skip-One] The Skip-One mechanism is claimed to lower round energy/latency while preserving long-term fairness; however, no formal analysis, convergence bound, or multi-round fairness metric (e.g., client participation distribution over epochs) is provided to substantiate that skipping at most one client per cluster does not introduce systematic bias in the aggregated model.
  3. [Proposed Mechanisms / random-k cross-aggregation] The random-k cross-aggregation is presented as enabling lightweight mixing without extending round duration, but the evaluation lacks an ablation on the choice of k, its effect on convergence rate, or a comparison of communication overhead versus full cross-cluster aggregation; this is load-bearing for the claim of energy efficiency without accuracy loss.
minor comments (2)
  1. [Abstract / Evaluation] The abstract and results sections refer to 'competitive accuracy' and 'faster convergence' without specifying the datasets (e.g., MNIST, CIFAR-10), number of satellites/clients, or exact baseline accuracies; these details should be added for reproducibility.
  2. [Throughout] Notation for key terms (LISL, GS, StarMask, etc.) should be defined consistently on first use and used uniformly throughout the manuscript.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We have revised the paper to address the concerns regarding the energy model details, added empirical fairness metrics for Skip-One, and included ablations for random-k cross-aggregation. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Evaluation / Energy Model] The central quantitative claims (over 100x GS communication reduction and 6x energy savings) rest entirely on the fidelity of the end-to-end energy model and Walker-Delta simulation described in the evaluation. The manuscript must provide the explicit equations and parameter values used for onboard computation energy (CPU/GPU power draw) and LISL communication energy, including all assumptions on link feasibility, interference, and hardware heterogeneity, as these directly determine the reported gains relative to baselines.

    Authors: We agree that explicit details are necessary to substantiate the claims. In the revised manuscript, Section 5.1 now includes the complete equations: onboard computation energy E_comp,i = P_cpu,i * T_train,i + P_gpu,i * T_infer,i (with P_cpu ranging 3-8W and P_gpu 10-20W for heterogeneous hardware); LISL energy E_lisl = P_tx * (d/c) * packet_count where P_tx=1.5W, assuming feasibility for distances under 1500km with narrow-beam lasers and negligible interference. All values, including Walker-Delta parameters and link probabilities, are tabulated in new Table 3. These additions directly support the 100x communication and 6x energy reductions. revision: yes

  2. Referee: [Proposed Mechanisms / Skip-One] The Skip-One mechanism is claimed to lower round energy/latency while preserving long-term fairness; however, no formal analysis, convergence bound, or multi-round fairness metric (e.g., client participation distribution over epochs) is provided to substantiate that skipping at most one client per cluster does not introduce systematic bias in the aggregated model.

    Authors: We acknowledge the absence of formal analysis in the original submission. The revised evaluation (Section 6.3) adds multi-round fairness metrics, including client participation histograms over 200 epochs showing average skip rate below 2% per client and participation variance under 4%, with no evidence of systematic bias. Skip-One is intended to maintain fairness in expectation by limiting skips to at most one per cluster. A rigorous convergence bound is not provided, as it would require substantial new theoretical development beyond this systems-focused work. revision: partial

  3. Referee: [Proposed Mechanisms / random-k cross-aggregation] The random-k cross-aggregation is presented as enabling lightweight mixing without extending round duration, but the evaluation lacks an ablation on the choice of k, its effect on convergence rate, or a comparison of communication overhead versus full cross-cluster aggregation; this is load-bearing for the claim of energy efficiency without accuracy loss.

    Authors: We appreciate this observation. The revised manuscript includes a new ablation study in Section 6.4 and Figure 8, testing k=1 to 5. For k=3, accuracy reaches 84% in 125 rounds (comparable to full aggregation at 120 rounds) while reducing inter-cluster communication by 60%. Overhead is quantified as O(k) links per round versus O(cluster count) for full mixing, preserving round duration and supporting the energy efficiency claims without accuracy degradation. revision: yes

standing simulated objections not resolved
  • Formal convergence bound or theoretical analysis for the Skip-One mechanism

Circularity Check

0 steps flagged

No significant circularity; performance claims are simulation outcomes

full rationale

The paper proposes CroSatFL with three mechanisms (StarMask, Skip-One, random-k cross-aggregation) and evaluates them via an end-to-end energy model on a Walker-Delta constellation simulation. No equations, derivations, or first-principles results are presented in the provided text that reduce any claimed prediction or result to fitted parameters or self-citations by construction. The quantitative claims (e.g., two orders of magnitude GS communication reduction, 6x energy savings) are explicitly simulation outputs relative to baselines, not tautological renamings or self-definitional loops. The framework's design choices are independent of the reported metrics, and no load-bearing self-citation chains or ansatz smuggling appear. This is a standard simulation-based systems paper whose central claims remain falsifiable against the model assumptions without internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities can be extracted. The energy model and Walker-Delta constellation are treated as given inputs from prior literature.

pith-pipeline@v0.9.0 · 5581 in / 1065 out tokens · 16080 ms · 2026-05-10T07:58:46.288864+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · 1 internal anchor

  1. [1]

    Satellite computing: Vision and challenges,

    S. Wang and Q. Li, “Satellite computing: Vision and challenges,”IEEE Internet of Things Journal, 2023

    work page 2023

  2. [2]

    Fedsn: A federated learning framework over heterogeneous leo satellite networks,

    Z. Lin, Z. Chen, Z. Fang, X. Chen, X. Wang, and Y . Gao, “Fedsn: A federated learning framework over heterogeneous leo satellite networks,” IEEE Transactions on Mobile Computing, 2025

    work page 2025

  3. [3]

    Fedleo: An offloading-assisted decentralized federated learning framework for low earth orbit satellite networks,

    Z. Zhai, Q. Wu, S. Yu, R. Li, F. Zhang, and X. Chen, “Fedleo: An offloading-assisted decentralized federated learning framework for low earth orbit satellite networks,”IEEE Transactions on Mobile Computing, 2024

    work page 2024

  4. [4]

    Fedorbit: Energy efficient federated learning for orbital edge computing using block minifloat arithmetic,

    M. R. Jabbarpour, B. Javadi, P. H. W. Leong, R. N. Calheiros, and D. Boland, “Fedorbit: Energy efficient federated learning for orbital edge computing using block minifloat arithmetic,”IEEE Transactions on Services Computing, 2024

    work page 2024

  5. [5]

    Advances in appfl: A comprehensive and extensible federated learning framework,

    Z. Li, S. He, Z. Yang, M. Ryu, K. Kim, and R. Madduri, “Advances in appfl: A comprehensive and extensible federated learning framework,” inCCGrid. IEEE, 2025

    work page 2025

  6. [6]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inAISTATS. PMLR, 2017

    work page 2017

  7. [7]

    Optimizing federated learning in leo satel- lite constellations via intra-plane model propagation and sink satellite scheduling,

    M. Elmahallawy and T. Luo, “Optimizing federated learning in leo satel- lite constellations via intra-plane model propagation and sink satellite scheduling,” inIEEE ICC. IEEE, 2023

    work page 2023

  8. [8]

    Edge selection and clustering for federated learning in optical inter-leo satellite constellation,

    C.-Y . Chen, L.-H. Shen, K.-T. Feng, L.-L. Yang, and J.-M. Wu, “Edge selection and clustering for federated learning in optical inter-leo satellite constellation,” inPIMRC. IEEE, 2023

    work page 2023

  9. [9]

    Enabling radiation tolerant heterogeneous gpu-based onboard data processing in space,

    F. C. Bruhn, N. Tsog, F. Kunkel, O. Flordal, and I. Troxel, “Enabling radiation tolerant heterogeneous gpu-based onboard data processing in space,”CEAS Space Journal, 2020

    work page 2020

  10. [10]

    Energy efficient client selection in federated learning for orbital edge computing,

    B. Al-Blewi, B. Javadi, and R. N. Calheiros, “Energy efficient client selection in federated learning for orbital edge computing,” inIEEE EDGE. IEEE, 2025

    work page 2025

  11. [11]

    A survey on inter-satellite links for low-earth orbit satellite networks,

    H. S. Cha, J. M. Kim, B. Lim, J. H. Lee, and Y . C. Ko, “A survey on inter-satellite links for low-earth orbit satellite networks,”Journal of Korean Institute of Communications and Information Sciences, 2022

    work page 2022

  12. [12]

    Towards space edge computing and onboard ai for real-time teleoperations,

    M. Gardill, W. Kinser, J. Budroweit, A. Al-Hourani, E. Dunkel, J. Swope, D. Evans, and M. Staebler, “Towards space edge computing and onboard ai for real-time teleoperations,” 2023

    work page 2023

  13. [13]

    Onboard processing in satellite communications using ai accelerators,

    F. Ortiz, V . Monzon Baeza, L. M. Garces-Socarras, J. A. V ´asquez- Peralvo, J. L. Gonzalez, G. Fontanesi, E. Lagunas, J. Querol, and S. Chatzinotas, “Onboard processing in satellite communications using ai accelerators,”Aerospace, 2023

    work page 2023

  14. [14]

    Eagleeye: Nanosatel- lite constellation design for high-coverage, high-resolution sensing,

    Z. Cheng, B. Denby, K. McCleary, and B. Lucia, “Eagleeye: Nanosatel- lite constellation design for high-coverage, high-resolution sensing,” in ASPLOS, 2024

    work page 2024

  15. [15]

    AI-SPACE: A cloud-edge aggregated artificial intelligent architecture for tiansuan constellation- assisted space-terrestrial integrated networks,

    X. Xu, Q. Wang, Y . Hou, and S. Wang, “AI-SPACE: A cloud-edge aggregated artificial intelligent architecture for tiansuan constellation- assisted space-terrestrial integrated networks,”IEEE Network, 2023

    work page 2023

  16. [16]

    Kodan: Addressing the computational bottleneck in space,

    B. Denby, K. Chintalapudi, R. Chandra, B. Lucia, and S. Noghabi, “Kodan: Addressing the computational bottleneck in space,” inASPLOS, 2023

    work page 2023

  17. [17]

    Towards efficient satellite computing through adaptive compression,

    C. Yanget al., “Towards efficient satellite computing through adaptive compression,”IEEE Transactions on Services Computing, 2024

    work page 2024

  18. [18]

    Satellite-based computing networks with federated learning,

    H. Chen, M. Xiao, and Z. Pang, “Satellite-based computing networks with federated learning,”IEEE Wireless Communications, 2022

    work page 2022

  19. [19]

    Federated learning in satellite constellations,

    B. Matthiesen, N. Razmi, I. Leyva-Mayorga, A. Dekorsy, and P. Popovski, “Federated learning in satellite constellations,”IEEE Net- work, 2024

    work page 2024

  20. [20]

    L2d2: Low latency distributed downlink for leo satellites,

    D. Vasisht, J. Shenoy, and R. Chandra, “L2d2: Low latency distributed downlink for leo satellites,” inSIGCOMM, 2021

    work page 2021

  21. [21]

    Battery-aware energy optimization for satellite edge computing,

    Q. Li, S. Wang, X. Ma, A. Zhou, Y . Wang, G. Huang, and X. Liu, “Battery-aware energy optimization for satellite edge computing,”IEEE Transactions on Services Computing, 2024

    work page 2024

  22. [22]

    Tiansuan constellation: Intelligent software-defined mi- crosatellite with orbital attention for sustainable development goals,

    Y . Tang, “Tiansuan constellation: Intelligent software-defined mi- crosatellite with orbital attention for sustainable development goals,” inICBDIC. Springer, 2022

    work page 2022

  23. [23]

    Asynchronous methods for deep reinforcement learning,

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep reinforcement learning,” inICML. PMLR, 2016

    work page 2016

  24. [24]

    Flower: A Friendly Federated Learning Research Framework

    D. J. Beutel, T. Topal, A. Mathur, X. Qiu, J. Fernandez-Marques, Y . Gao, L. Sani, K. H. Li, T. Parcollet, P. P. B. De Gusm ˜aoet al., “Flower: A friendly federated learning research framework,”arXiv preprint arXiv:2007.14390, 2020

    work page internal anchor Pith review arXiv 2007

  25. [25]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016

    work page 2016

  26. [26]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, 2002

    work page 2002

  27. [27]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

    work page 2009

  28. [28]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classi- fication,

    P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classi- fication,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019

    work page 2019