Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

A Semi-Supervised Federated Learning Framework with Hierarchical Clustering Aggregation for Heterogeneous Satellite Networks

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a semi-supervised federated learning framework with hierarchical clustering aggregation can reduce LEO satellite training time up to 3x and energy consumption up to 4x without losing accuracy.

desk verdict The 3x speedup claim is undermined by an internal inconsistency in the processing-time formula, so the paper needs a major revision before the headline efficiency results can be trusted. read the letter →

arxiv 2507.22339 v1 pith:PB7MQVHR submitted 2025-07-30 cs.DC cs.LG

Zhuocheng Liu , Zhishu Shen , Qiushi Zheng , Tiehua Zhang , Zheng Lei , Jiong Jin This is my paper · ORCID
classification cs.DCcs.LG
keywords semi-supervisedfederatedlearningLEOsatellitenetworkshierarchicalclusteringaggregationsparsificationadaptivequantizationnon-IIDdataenergyefficiencyprocessingtime
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes SFedSat, a semi-supervised federated learning framework for low-Earth-orbit (LEO) satellite constellations where most satellites have no labeled data. It argues that by clustering satellites into groups based on both their data gradients and their geographic positions, aggregating model updates first within each cluster and then at ground stations, and compressing every update through sparsification and adaptive quantization, the training process can be made much faster and cheaper. The claimed result is that SFedSat reaches the same target accuracy in about a third of the processing time and about a quarter of the energy of the compared centralized federated baselines, while also handling non-IID and partially unlabeled data.

What carries the argument

The load-bearing mechanism is the hierarchical two-stage aggregation built on joint-feature clustering. Satellites are grouped by K-means on joint feature vectors that concatenate normalized gradient cosine similarity and normalized geographic distance; each cluster picks a parameter-server satellite that aggregates local updates, and ground stations perform the final global aggregation. Semi-supervised training alternates supervised loss on ground-station data with pseudo-labeled FixMatch and CutMix losses on unlabeled satellites, and a staleness-aware weight (inverse of staleness) selects the fastest fraction of clients. Communication is compressed by random sparsification with unbiased scaling and adaptive stochastic quantization to 4 or 8 bits depending on gradient change; the unbiasedness of Q ensures no systematic bias in aggregation.

What would settle it

Run SFedSat against C-FedAvg on a satellite network simulator that models intermittent inter-satellite visibility, Doppler shifts, and topology changes; if the wall-clock processing time and total energy consumption no longer show roughly 3x and 4x reductions at equal accuracy, the central claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that semi-supervised federated learning, previously tested mainly in stable networks, can be made practical for heterogeneous LEO satellite networks through a two-stage hierarchical aggregation design. In SFedSat, ground stations hold the only labeled data and run supervised updates; parameter-server satellites within each cluster propagate these models to fully unlabeled satellites, which train with pseudo-labels, FixMatch confidence filtering, and CutMix augmentation. A joint feature vector, built from gradient cosine similarity and geographic distance, feeds K-means clustering, and a staleness-aware semi-asynchronous mechanism aggregates only the fastest fraction of clients per round. Sparsification plus adaptive 4/8-bit unbiased quantization shrinks transmitted updates by roughly 5-7x. The paper reports that on non-IID CIFAR-10 and SAT-6 satellite imagery, this setup reaches 80% target accuracy with up to 3x less processing time and up to 4x less energy than the compared methods, and it gives a convergence bound with error $O(1/(\eta m) + \iota^2 + \lambda \varepsilon_{ssl} + \omega)$.

Load-bearing premise

The claimed speedups and energy savings assume an idealized satellite link model with constant bandwidth and noise and no intermittent disconnections; if real orbital communication dynamics are less stable, the 3x/4x gains may not transfer to practice.

Editorial extensions

If this is right

  • LEO constellations could run onboard federated learning with far lower energy budgets, making distributed model training a realistic onboard task rather than a ground-station-only operation.
  • Only ground stations need labeled data; satellites can train on raw, unlabeled imagery through pseudo-labeling, which matches real-world annotation scarcity.
  • Clustering by joint gradient and geographic features mitigates non-IID data, so accuracy is maintained even with strongly skewed per-client class distributions.
  • Compression of 5-7x in transmitted model size means communication windows, which are short and intermittent in satellite networks, become less of a bottleneck.
  • The 3x/4x processing-time and energy reductions, if realized, would translate directly into lower operating costs and more frequent model updates for Earth-observation and remote-sensing applications.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to replace the fixed Shannon-rate channel model with orbital-mechanics-aware link schedules (intermittent visibility, Doppler, outages); the claimed 3x/4x gains would face a stricter test.
  • The same joint-feature clustering and two-stage aggregation could transfer to satellite-terrestrial or high-altitude-platform networks, where the 'geographic distance' component has a direct physical meaning.
  • Because the convergence proof in Section IV-E uses a simplified synchronous update without staleness weights or the semi-supervised loss, a formal convergence guarantee for the actual asynchronous algorithm remains open; closing that gap would strengthen the framework's theoretical basis.
  • The adaptive 4/8-bit quantization could be turned into an online policy that learns the gradient-change threshold Z per cluster, rather than using a fixed global value.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes SFedSat, a semi-supervised federated learning framework for LEO satellite networks structured around hierarchical clustering aggregation. The core ideas are: clustering satellites into K clusters based on a joint feature vector combining gradient cosine similarity and geographic proximity; a two-stage aggregation scheme with satellite Parameter Servers (PSs) and Ground Stations (GSs); a staleness-aware semi-asynchronous intra-cluster update; and a sparsification plus adaptive weight quantization mechanism. The authors claim that SFedSat reduces processing time by up to 3x and energy consumption by up to 4x relative to baselines while maintaining model accuracy, and they provide a convergence analysis with an O(1/(eta m) + iota^2 + lambda*epsilon_ssl + omega) error bound. Experiments are carried out on a simulated satellite network testbed using CIFAR-10 (100 satellites) and SAT-6 (3000 satellites) against four baselines (C-FedAvg, H-BASE, FedMatch, SemiFL).

Significance. If the efficiency claims are supported, SFedSat would be a meaningful contribution to resource-constrained federated learning in satellite networks, where communication windows and energy budgets are tight. The bundle of techniques - gradient-guided clustering, client selection based on completion speed, and unbiased compression - is plausible and worth investigating. However, the central processing-time model in Eq. (8) is internally inconsistent with the paper's own parallel-cluster description, the convergence proof abstracts away the distinctive algorithmic elements, and the experiments currently lack the statistical and ablation evidence needed to substantiate the headline quantitative gains. The paper is therefore not yet ready for publication in its present form, but the proposed framework is promising and the issues identified appear addressable within the scope of a major revision.

major comments (4)
  1. [Sec. III-C, Eq. (8)] The text states that satellite clusters engage in parallel local training and that the ground-station delay is 'primarily determined by the slowest cluster among the set s_k', yet Eq. (8) defines Tc = Σ_{s_k∈gKn} (max_{i∈s_k} T_i^m + T_{s_k} + t_broc). This sum makes the per-round processing time grow linearly with the number of clusters, which contradicts the stated parallel schedule. Since the abstract's 'up to 3x processing time' claim is evaluated 'as defined in Equation 8', the headline efficiency result is not currently supported: if the implementation follows the sum, the parallel-execution claim is false; if it follows the slowest-cluster description, the published formula is wrong. The authors must correct the model and re-evaluate the processing-time comparisons in Fig. 7.
  2. [Sec. IV-E, Eq. (43) and Eq. (51)] The convergence analysis is conducted for the simplified synchronous update in Eq. (43), which omits the actual aggregation rule in Eq. (27): the staleness weights p_i^m, the client-selection fraction epsilon, and the semi-supervised loss terms L_fix and L_cutmix are all absent from the proof. Moreover, the final error bound in Eq. (51) introduces a term lambda*epsilon_ssl without a prior definition or derivation; the preceding inequalities (46)-(50) contain no lambda or epsilon_ssl. Thus the stated convergence guarantee does not apply to the algorithm as described. Please either extend the proof to the actual update rule, or clearly state which simplified variant is being analyzed and remove the unsupported epsilon_ssl term.
  3. [Sec. V-C and Table II] The experimental methodology does not currently support the empirical claims. Specifically, the cluster count K is selected after inspecting the test-accuracy curves in Fig. 3, and the same test set appears to be used both for model selection and final comparison; no error bars, random seeds, or repeated runs are reported. In addition, Table II reports only total transmitted weight sizes before/after compression and does not include any accuracy figures, so the statement in Sec. V-C that compression 'maintains or even improves accuracy' is not backed by the displayed evidence. Please add multiple-seed statistics, a principled validation procedure for K, and an accuracy column (or an accuracy-vs-compression plot) in the compression ablation.
  4. [Algorithm 1, lines 5-14] In Algorithm 1, the supervised-learning block (lines 5-14) appears inside the parallel client loop (line 4), but the surrounding text and Sec. IV-C state that supervised learning is performed at the ground stations. As written, every satellite would construct the labeled dataset D_s and update using L_s, which contradicts the experimental setup where all labeled data is located at GSs and satellite clients are entirely unlabeled. Please restructure the pseudocode so that the GS-level supervised update is performed outside the satellite-client loop, matching the described alternating training strategy.
minor comments (6)
  1. [Sec. III-B and Eq. (7)] The symbol D_i is used both for the local dataset in Eq. (1) and for the data volume in the computation-time formula t_cmp^i = D_i Q_i / f_i in Eq. (7); please distinguish dataset size in samples from data volume in bits to avoid ambiguity.
  2. [Eq. (17)] The expression 'θ · Hcos ∥ (1 − θ) · Hgeo' uses parallel bars to indicate concatenation; please define this notation explicitly, as it can be mistaken for a norm.
  3. [Eq. (33) and Sec. V-A] The text in Sec. V-A says global models with gradient magnitudes below 0.01 are quantized to 4 bits, whereas Eq. (33) sets b_i^m = 4 when Δdiff(m) ≤ Z; please reconcile the direction of the threshold condition.
  4. [Sec. IV-E, Eq. (46)] The inequality E[∥\barΔ_m∥^2] + E[∥\barμ_m∥^2] ≤ 2ι² + 2ω does not follow directly from Assumptions 2-4 as written, because Assumption 4 bounds E[∥μ_i(m)∥²] in terms of ∥Δw_i^m∥², not by a constant; please state the additional moment assumptions used to obtain this bound.
  5. [References] References [19], [28], and [31] contain the typo 'Proceddings' and should read 'Proceedings'.
  6. [Sec. V-C, Fig. 7 discussion] The sentence 'the completion time of C-FedAvg is 2 × H-BASE's and 3× SFedSat's' is ambiguous about whether the multiplication applies to the time values or to the ratios; please rephrase for clarity.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity found; efficiency claims are empirical, while Eq. (8) and Eq. (51) raise support/consistency issues that are correctness gaps rather than circular reductions.

full rationale

The headline processing-time and energy claims are empirical results from a simulated satellite testbed compared against external baselines (C-FedAvg, H-BASE, FedMatch, SemiFL) on CIFAR-10 and SAT-6; they are measured outcomes, not fitted parameters renamed as predictions. The clustering mechanism (Eqs. 13-17) uses cosine similarity of client gradients and geographic distance as clustering features, which is self-referential only in the benign sense that gradients summarize local data; it does not define the claimed accuracy, time, or energy results in terms of themselves. The convergence analysis in Sec. IV-E is under-supported: Eq. (43) explicitly adopts a simplified synchronous update that omits the staleness weights and semi-supervised losses used in the actual update (Eqs. 24-27), and Eq. (51) introduces lambda*epsilon_ssl without derivation, with epsilon_ssl merely named as 'the unsupervised estimation error.' This makes the convergence bound vacuous as a proof for SFedSat, but it is a missing-derivation/correctness issue, not a circular equivalence. Similarly, Sec. III-C states that the GS delay is determined by the slowest cluster, while Eq. (8) sums per-cluster times; this internal inconsistency could invalidate the 3x processing-time comparison if Eq. (8) was used in evaluation, but it is a model-formulation flaw rather than a reduction of a prediction to its inputs. The paper's self-citations ([2], [7]) are background references and are not load-bearing for any claimed derivation. Overall, no step in the claimed derivation chain is equivalent to its own input by construction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central efficiency claims rest on several hand-set hyperparameters (theta, K, Z, epsilon, tau, mu, lambda) and on assumptions about the convergence proof's applicability, the reliability of gradient-based clustering, and the fidelity of the satellite simulator. None of the assumptions has independent evidence outside the paper's own simulated experiments.

free parameters (7)
  • theta (joint feature weight) = 0.4
    Balances gradient similarity and geographic proximity in Eq. 17; hand-set and central to the clustering mechanism.
  • K (number of clusters) = 6 for CIFAR-10, 50 for SAT-6
    Selected from test accuracy curves in Fig. 3, then used in all later efficiency comparisons; post hoc tuning on the test set.
  • Z (gradient-change threshold) = 0.01
    Controls the switch between 8-bit and 4-bit quantization in Eq. 33; hand-set.
  • epsilon (client selection rate) = 0.6
    Fraction of fastest clients selected for aggregation; hand-set in Table I.
  • tau (pseudo-label confidence threshold) = not specified
    Used to construct the high-confidence pseudo-labeled set in Eq. 20; no value given in Algorithm 1 or Table I.
  • mu (CutMix Beta parameter) = not specified
    Controls the mixing ratio in Eq. 21; no value given.
  • lambda (loss weighting) = not specified
    Weights FixMatch vs CutMix loss in Eq. 24; no value given.
assumptions (4)
  • domain assumption Assumptions 1-4 hold (L-smoothness, bounded stochastic gradient variance, bounded client drift, unbiased compression with bounded relative variance).
    Used in Sec. IV-E to derive the convergence bound; these are standard but unverified for the satellite FL setting.
  • ad hoc to paper The simplified synchronous update in Eq. 43 represents the actual SFedSat algorithm in Eq. 27.
    The proof does not incorporate staleness weights, client selection fraction, clustering, or semi-supervised losses; the paper asserts the result carries over without derivation.
  • domain assumption Cosine similarity of gradient updates is a valid proxy for data-distribution similarity among satellite clients.
    Invoked in Sec. IV-B (Eq. 13-14) to justify the joint feature clustering; no empirical or theoretical validation is given beyond the main experiments.
  • domain assumption The simulated testbed (fixed orbit parameters, Shannon-rate model, timing and energy models in Sec. III-C) is representative of real LEO satellite networks.
    All quantitative claims about processing time and energy savings are produced by this simulator; real intermittent visibility and topology dynamics are not modeled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Semi-Supervised Federated Learning Framework with Hierarchical Clustering Aggregation for Heterogeneous Satellite Networks." pith.science (2026). https://pith.science/paper/PB7MQVHR

@misc{pith2026250722339,
  author       = {Pith},
  title        = {Pith review of: A Semi-Supervised Federated Learning Framework with Hierarchical Clustering Aggregation for Heterogeneous Satellite Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PB7MQVHR}},
  note         = {Machine review of arXiv:2507.22339}
}
read the original abstract

Low Earth Orbit (LEO) satellites are emerging as key components of 6G networks, with many already deployed to support large-scale Earth observation and sensing related tasks. Federated Learning (FL) presents a promising paradigm for enabling distributed intelligence in these resource-constrained and dynamic environments. However, achieving reliable convergence, while minimizing both processing time and energy consumption, remains a substantial challenge, particularly in heterogeneous and partially unlabeled satellite networks. To address this challenge, we propose a novel semi-supervised federated learning framework tailored for LEO satellite networks with hierarchical clustering aggregation. To further reduce communication overhead, we integrate sparsification and adaptive weight quantization techniques. In addition, we divide the FL clustering into two stages: satellite cluster aggregation stage and Ground Stations (GSs) aggregation stage. The supervised learning at GSs guides selected Parameter Server (PS) satellites, which in turn support fully unlabeled satellites during the federated training process. Extensive experiments conducted on a satellite network testbed demonstrate that our proposal can significantly reduce processing time (up to 3x) and energy consumption (up to 4x) compared to other comparative methods while maintaining model accuracy.

Figures

Figures reproduced from arXiv: 2507.22339 by the authors.

Figure 1
Figure 1. The satellite network model. To mitigate the impact of system heterogeneity, current mainstream research focuses on staleness-aware weighted ag￾gregation strategies, which adjust each client’s contribution to the global model based on the timeliness of its updates. For example, You et al. proposed a time-decay-based weighting mechanism that assigns higher weights to more recent updates while reducing the influence o… view at source ↗
Figure 2
Figure 2. Overview of our proposed framework SFedSat. feature vectors dynamic clustering algorithm (line 1 in Algo￾rithm 1). The objective is to group satellite clients that share both similar data characteristics and geographical proximity, as such clients are more likely to generate model updates with aligned gradient directions. To evaluate the similarity of data distributions among satellites, we employ cosine similarity … view at source ↗
Figure 6
Figure 6. Total energy consumption for different methods. (a) CIFAR-10 (b) SAT-6 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figures from the paper (2 more)
Figure 7
Figure 7. Figure 7: Processing time for different methods. FL processing time and energy consumption: In subse￾quent experiments, all algorithms are evaluated under the con￾dition of reaching a target accuracy of 80%. The FL processing time includes the computation time and communication …
Figure 5
Figure 5. Figure 5: Loss performance for different methods. Training convergence: To validate the effectiveness of our proposed framework, we compare the global model accuracy and training loss over 500 communication rounds against four baselines on different datasets. This evaluation aim…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FedRings: A Scalable and Topology-Aware Federated Learning Framework for LEO Satellite Constellations

    cs.DC 2026-08 conditional novelty 5.0 of 10

    FedRings arranges LEO satellites into ring structures with predictive, sparsified model-update propagation, claiming improved communication efficiency and convergence in simulation.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Revolutionizing future connectivity: A contem- porary survey on AI-empowered satellite-based non-terrestrial networks in 6G,

    S. Mahboob and L. Liu, “Revolutionizing future connectivity: A contem- porary survey on AI-empowered satellite-based non-terrestrial networks in 6G,” IEEE Communications Surveys & Tutorials , vol. 26, no. 2, pp. 1279–1321, 2024

    2024

  2. [2]

    A survey of next-generation computing technologies in space-air-ground integrated networks,

    Z. Shen et al. , “A survey of next-generation computing technologies in space-air-ground integrated networks,” ACM Computing Surveys , vol. 56, no. 1, 2023

    2023

  3. [3]

    Energy-efficient computation peer offloading in satellite edge computing networks,

    X. Zhang, J. Liu, R. Zhang, Y . Huang, J. Tong, N. Xin, L. Liu, and Z. Xiong, “Energy-efficient computation peer offloading in satellite edge computing networks,” IEEE Transactions on Mobile Computing, vol. 23, no. 4, pp. 3077–3091, 2024

    work page 2024

  4. [4]

    Satellite edge intelligence: DRL-based resource management for task inference in LEO-based satellite-ground collaborative networks,

    W. Fan, Q. Meng, G. Wang, H. Bian, Y . Liu, and Y . Liu, “Satellite edge intelligence: DRL-based resource management for task inference in LEO-based satellite-ground collaborative networks,” IEEE Transactions on Mobile Computing , pp. 1–18, 2025

    work page 2025

  5. [5]

    Satellite internet of things for smart agriculture applications: A case study of computer vision,

    J. Liu, W. Jiang, H. Han, M. He, and W. Gu, “Satellite internet of things for smart agriculture applications: A case study of computer vision,” in Proceedings of the IEEE International Conference on Sensing, Communication, and Networking (SECON) , pp. 66–71, 2023

    work page 2023

  6. [6]

    Semi-supervised federated learning for assessing building damage from satellite imagery,

    Y . Zhang, Y . Gong, and Y . Guo, “Semi-supervised federated learning for assessing building damage from satellite imagery,” in Proceedings of the IEEE International Conference on Communications (ICC) , pp. 3821– 3826, 2024

    work page 2024

  7. [7]

    APT- SAT: An adaptive DNN partitioning and task offloading framework within collaborative satellite computing environments,

    S. Peng, Z. Shen, Q. Zheng, X. Hou, D. Jiang, J. Yuan, and J. Jin, “APT- SAT: An adaptive DNN partitioning and task offloading framework within collaborative satellite computing environments,” IEEE Transac- tions on Network Science and Engineering , pp. 1–14, 2025

    work page 2025

  8. [8]

    Adaptive configuration for heterogeneous participants in decentralized federated learning,

    Y . Liao, Y . Xu, H. Xu, L. Wang, and C. Qian, “Adaptive configuration for heterogeneous participants in decentralized federated learning,” in Proceedings of the IEEE Conference on Computer Communications (INFOCOM), pp. 1–10, 2023

    work page 2023

Show all 43 references
  1. [9]

    Resource management for MEC assisted multi-layer federated learning framework,

    H. Li, Y . Pan, H. Zhu, P. Gong, and J. Wang, “Resource management for MEC assisted multi-layer federated learning framework,” IEEE Transactions on Wireless Communications , vol. 23, no. 6, pp. 5680– 5693, 2024

    2024

  2. [10]

    Communication-efficient satellite-ground federated learning through progressive weight quantization,

    C. Yang, J. Yuan, Y . Wu, Q. Sun, A. Zhou, S. Wang, and M. Xu, “Communication-efficient satellite-ground federated learning through progressive weight quantization,” IEEE Transactions on Mobile Com- puting, vol. 23, no. 9, pp. 8999–9011, 2024

    2024

  3. [11]

    FedSN: A fed- erated learning framework over heterogeneous LEO satellite networks,

    Z. Lin, Z. Chen, Z. Fang, X. Chen, X. Wang, and Y . Gao, “FedSN: A fed- erated learning framework over heterogeneous LEO satellite networks,” IEEE Transactions on Mobile Computing, vol. 24, no. 3, pp. 1293–1307, 2025

    2025

  4. [12]

    ALANINE: A novel decentralized personalized federated learning for heterogeneous leo satellite constellation,

    L. Zhao, S. Geng, X. Tang, A. Hawbani, Y . Sun, L. Xu, and D. Tarchi, “ALANINE: A novel decentralized personalized federated learning for heterogeneous leo satellite constellation,” IEEE Transactions on Mobile Computing, vol. 24, no. 8, pp. 6945–6960, 2025

    2025

  5. [13]

    Federated learning on non-IID data silos: An experimental study,

    Q. Li, Y . Diao, Q. Chen, and B. He, “Federated learning on non-IID data silos: An experimental study,” in Proceedings of the IEEE International Conference on Data Engineering (ICDE) , pp. 965–978, 2022

    2022

  6. [14]

    Self-supervised spatio-temporal representation learning of satellite image time series,

    I. Dumeur, S. Valero, and J. Inglada, “Self-supervised spatio-temporal representation learning of satellite image time series,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 17, pp. 4350–4367, 2024

    2024

  7. [15]

    Energy-efficient federated learning for earth observation in LEO satellite systems,

    T. Xiong, X. Xu, P. Du, Y . Zhang, R. Zhang, and J. Liu, “Energy-efficient federated learning for earth observation in LEO satellite systems,” in Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC), pp. 01–06, 2024

    2024

  8. [16]

    Connection- density-aware satellite-ground federated learning via asynchronous dy- namic aggregation,

    Z. Xu, M. Jin, J. Lin, Y . Liu, J. Xu, Z. Xiong, and H. Cai, “Connection- density-aware satellite-ground federated learning via asynchronous dy- namic aggregation,” Future Generation Computer Systems , vol. 155, pp. 312–323, 2024

    2024

  9. [17]

    Communication-efficient federated learning for LEO constellations integrated with HAPs using hybrid NOMA-OFDM,

    M. Elmahallawy, T. Luo, and K. Ramadan, “Communication-efficient federated learning for LEO constellations integrated with HAPs using hybrid NOMA-OFDM,” IEEE Journal on Selected Areas in Communi- cations, vol. 42, no. 5, pp. 1097–1114, 2024

    2024

  10. [18]

    Energy-efficient resource manage- ment for federated learning in LEO satellite IoT,

    A. Zhou, Y . Wang, and Q. Zhang, “Energy-efficient resource manage- ment for federated learning in LEO satellite IoT,” in Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC) , pp. 1–6, 2024

    2024

  11. [19]

    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,” in Proceddings of the Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC),...

    2023

  12. [20]

    A survey on satellite networks with federated learning to analyze data or manage resource,

    J. Oh, D. Lee, T. P. Truong, D. Hur, S. Hong, and S. Cho, “A survey on satellite networks with federated learning to analyze data or manage resource,” in Proceedings of the International Conference on Artificial 12 Intelligence in Information and Communication (ICAIIC) , pp. 0...

    2025

  13. [21]

    Decomposition and meta-DRL based multi-objective optimization for asynchronous federated learning in 6G-satellite systems,

    Y . Zhou, L. Lei, X. Zhao, L. You, Y . Sun, and S. Chatzinotas, “Decomposition and meta-DRL based multi-objective optimization for asynchronous federated learning in 6G-satellite systems,” IEEE Journal on Selected Areas in Communications , vol. 42, no. 5, pp. 1115–1129, 2024

    2024

  14. [22]

    Cross-domain federated computation offloading for age of information minimization in satellite-airborne-terrestrial networks,

    X. Xia, H. H. Esmat, K. Dyer, B. Lorenzo, and L. Guo, “Cross-domain federated computation offloading for age of information minimization in satellite-airborne-terrestrial networks,” in Proceedings of the IEEE Annual International Symposium on Personal, Indoor and Mobile Radio ...

    2023

  15. [23]

    Exploitation maximization of unlabeled data for federated semi-supervised learning,

    S. Chen and J. Shen, “Exploitation maximization of unlabeled data for federated semi-supervised learning,” IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 9, no. 2, pp. 2039–2044, 2025

    2025

  16. [24]

    Federated semi- supervised learning with inter-client consistency & disjoint learning,

    W. Jeong, J. Yoon, E. Yang, and S. J. Hwang, “Federated semi- supervised learning with inter-client consistency & disjoint learning,” in Proceedings of the International Conference on Learning Represen- tations (ICLR), 2021

    2021

  17. [25]

    SemiFL: Semi-supervised federated learning for unlabeled clients with alternate training,

    E. Diao, J. Ding, and V . Tarokh, “SemiFL: Semi-supervised federated learning for unlabeled clients with alternate training,” in Proceedings of the Advances in Neural Information Processing Systems (NIPS), vol. 35, pp. 17871–17884, 2022

    2022

  18. [26]

    Toward fast personalized semi-supervised federated learning in edge networks: Algorithm design and theoretical guarantee,

    S. Wang, Y . Xu, Y . Yuan, and T. Q. S. Quek, “Toward fast personalized semi-supervised federated learning in edge networks: Algorithm design and theoretical guarantee,” IEEE Transactions on Wireless Communica- tions, vol. 23, no. 2, pp. 1170–1183, 2024

    2024

  19. [27]

    Boosting semi-supervised federated learning by effectively exploiting server-side knowledge and client-side unconfident samples,

    H. Liu, Y . Mi, Y . Tang, J. Guan, and S. Zhou, “Boosting semi-supervised federated learning by effectively exploiting server-side knowledge and client-side unconfident samples,” Neural Networks, vol. 188, p. 107440, 2025

    2025

  20. [28]

    Hybrid-FL for wireless networks: Cooperative learning mechanism us- ing non-IID data,

    N. Yoshida, T. Nishio, M. Morikura, K. Yamamoto, and R. Yonetani, “Hybrid-FL for wireless networks: Cooperative learning mechanism us- ing non-IID data,” in Proceddings of the IEEE International Conference on Communications (ICC) , pp. 1–7, 2020

    2020

  21. [29]

    Gradient scheduling with global momentum for asynchronous federated learning in edge environment,

    H. Wang, R. Li, C. Li, P. Zhou, Y . Li, W. Xu, and S. Guo, “Gradient scheduling with global momentum for asynchronous federated learning in edge environment,” IEEE Internet of Things Journal , vol. 9, no. 19, pp. 18817–18828, 2022

    2022

  22. [30]

    A triple-step asynchronous fed- erated learning mechanism for client activation, interaction optimization, and aggregation enhancement,

    L. You, S. Liu, Y . Chang, and C. Yuen, “A triple-step asynchronous fed- erated learning mechanism for client activation, interaction optimization, and aggregation enhancement,” IEEE Internet of Things Journal , vol. 9, no. 23, pp. 24199–24211, 2022

    2022

  23. [31]

    Towards efficient asynchronous federated learning in heterogeneous edge environments,

    Y . Zhou, X. Pang, Z. Wang, J. Hu, P. Sun, and K. Ren, “Towards efficient asynchronous federated learning in heterogeneous edge environments,” in Proceedings of the IEEE Conference on Computer Communications (ICC), pp. 2448–2457, 2024

    2024

  24. [32]

    FixMatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “FixMatch: Simplifying semi- supervised learning with consistency and confidence,” in Proceedings of the Advances in Neural Information Processing Systems (NIPS), vol. 3...

    2020

  25. [33]

    RandAugment: Practical automated data augmentation with a reduced search space,

    E. D. Cubuk, B. Zoph, J. Shlens, and Q. Le, “RandAugment: Practical automated data augmentation with a reduced search space,” in Proceed- ings of the Advances in Neural Information Processing Systems (NIPS) , vol. 33, pp. 18613–18624, 2020

    2020

  26. [34]

    AC-SGD: Adaptively compressed sgd for communication-efficient distributed learning,

    G. Yan, T. Li, S.-L. Huang, T. Lan, and L. Song, “AC-SGD: Adaptively compressed sgd for communication-efficient distributed learning,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 9, pp. 2678– 2693, 2022

    2022

  27. [35]

    On the conver- gence of fedavg on non-iid data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the conver- gence of fedavg on non-iid data,” in Proceedings of the International Conference on Learning Representations (ICLR) , 2020

    2020

  28. [36]

    Deepsat: a learning framework for satellite imagery,

    S. Basu, S. Ganguly, S. Mukhopadhyay, R. DiBiano, M. Karki, and R. Nemani, “Deepsat: a learning framework for satellite imagery,” in Proceedings of the International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL) , 2015

    2015

  29. [37]

    Wide residual networks.,

    S. Zagoruyko and N. Komodakis., “Wide residual networks.,” in arXiv preprint arXiv:1605.07146, 2016

    2016 arXiv

  30. [38]

    QSGD: Communication-efficient SGD via gradient quantization and encoding,

    D. Alistarh, D. Grubic, J. Li, R. Tomioka, and M. V ojnovic, “QSGD: Communication-efficient SGD via gradient quantization and encoding,” in Proceedings of the International Conference on Neural Information Processing Systems (NIPS) , vol. 30, pp. 1707–1718, 2017

    2017

  31. [39]

    Delay optimization for cooperative multi-tier computing in integrated satellite-terrestrial networks,

    X. Zhu and C. Jiang, “Delay optimization for cooperative multi-tier computing in integrated satellite-terrestrial networks,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 2, pp. 366–380, 2023

    2023

  32. [40]

    Satellite edge computing with collaborative computation offloading: An intelligent deep determin- istic policy gradient approach,

    H. Zhang, R. Liu, A. Kaushik, and X. Gao, “Satellite edge computing with collaborative computation offloading: An intelligent deep determin- istic policy gradient approach,” IEEE Internet of Things Journal, vol. 10, no. 10, pp. 9092–9107, 2023

    2023

  33. [41]

    Client-edge-cloud hier- archical federated learning,

    L. Liu, J. Zhang, S. Song, and K. B. Letaief, “Client-edge-cloud hier- archical federated learning,” in Proceedings of the IEEE International Conference on Communications (ICC) , pp. 1–6, 2020. Zuocheng Liu received his B.E. degree from the School of Computer Science and Artif...

    2020

  34. [2006]

    A distinctive feature of his research is its real- world impact and industry focus

    He has research interests in aviation business models, aviation market analysis, aviation regulation and policy analysis, and uncrewed aircraft systems. A distinctive feature of his research is its real- world impact and industry focus. He is also a Board member of Aviation Ae...

    2006

  35. [2024]

    His research interests encompass col- laborative learning/optimization, edge intelligence, graph learning, and the Internet of Things

    He is currently an Assistant Professor with the School of Computer Science and Technology, Tongji University. His research interests encompass col- laborative learning/optimization, edge intelligence, graph learning, and the Internet of Things. Zheng Lei is the inaugural Profe...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.