Pith. sign in

REVIEW 4 major objections 4 minor 38 references

This paper claims that under domain shift in federated graph learning, most gradient coordinates are dominated by domain-specific noise, and a server-side top-ρ mask keeping roughly 5-10% of coordinates—plus influence-regularised weighting—

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 16:26 UTC pith:MWRTCMNO

load-bearing objection A practical server-side gradient-mask idea with a broad empirical study, but the convergence proof is unsupported and the paper contradicts itself on its own algorithm. the 4 major comments →

arxiv 2509.18171 v5 pith:MWRTCMNO submitted 2025-09-17 cs.LG

Dissecting Federated-Graph Aggregation under Domain Shift: Importance-Aware Aggregation via Empirical Analysis

Zhanting Zhou , Kahou Tam , Zeyu Ma , Ziqiang Zheng This is my paper
classification cs.LG
keywords federated graph learningdomain shiftgradient sparsificationimportance-aware aggregationserver-side maskinginfluence-regularised momentumnode classificationdomain skew
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that aggregation is not the real problem in federated graph learning under domain shift; the gradient signal itself is. It claims that as clients optimize on different graph domains, the majority of gradient coordinates become dominated by domain-specific variance, so standard averaging dilutes the few coordinates that matter. The proposed fix, FedIA, is a server-side two-stage process: a global top-ρ mask keeps only the most informative ~5-10% of coordinates, and an influence-regularised momentum weight down-weights outlier clients. If correct, this would mean a zero-communication, server-only projection is enough to stabilize FGL and improve accuracy by up to 11.58 percentage points across two graph datasets and two backbones. A convergence sketch argues the projected update still achieves the optimal O(σ²/√T) rate under slow mask drift.

Core claim

The paper's central discovery is that the bottleneck in domain-skewed federated graph learning is not the weighting rule but the noisy gradient signal: a large fraction (potentially over 90%) of gradient dimensions are dominated by domain-specific variance. Standard averaging then attenuates the few coordinates that carry genuine learning signal, producing an aggregation-induced signal dilution. To counter this, FedIA preconditions gradients before aggregation: Stage 1 computes a server-side global importance mask by averaging absolute client gradients and retaining only the top-ρ fraction of coordinates; Stage 2 reweights clients by exponentially decaying contributions whose masked gradient

What carries the argument

The central object is the global importance mask M^t = TopRatioMask((1/|S_t|) Σ_k |g_k^t|, ρ): a binary mask over gradient coordinates, recomputed each round from the mean absolute client gradient, that keeps the top-ρ fraction. It does the work of projecting every client update onto a common sparse subspace before aggregation, so coordinates that only a few domains strongly express are not attenuated by near-zero contributions. The second piece is the influence-regularised momentum weight α_k^{t+1} = α_k^t exp(-λ ||g_{k,M^{t+1}}^{t+1} - \bar{g}_{M^{t+1}}||_2), which down-weights clients whose masked gradients are far from the aggregated mean; it acts as a stabilizer when the remaining spars

Load-bearing premise

The whole guarantee rests on the assumption that the set of gradient coordinates the mask selects stays nearly the same from round to round; the paper only measures this on its own runs and does not include the correction mechanism the cited convergence theorem depends on.

What would settle it

Replace the importance mask with a fresh random top-ρ mask each round; if accuracy and stability do not collapse, then the specific coordinate selection is not the load-bearing mechanism. Or measure IOU between consecutive masks and check whether any divergence run coincides with drift exceeding ρ.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • FedIA, applied on top of FedAvg or other aggregators, improves accuracy by up to 11.58 percentage points on Twitch Gamers and Wikipedia Network under cross-domain and domain-skew settings.
  • Because the mask and weighting live entirely on the server and reuse the standard gradient payload, the method adds no extra uplink traffic and negligible server memory, making it a drop-in module.
  • Methods that fail under domain skew, such as FedDyn dropping to random-chance accuracy, are restored to competitive performance once FedIA's projection and weighting are applied.
  • The ablation shows the projection stage accounts for most of the gain, while the momentum stage prevents numerical collapse under extreme skew, so a practitioner needs both stages.
  • The convergence sketch claims the projected update retains the O(σ²/√T) rate for non-convex smooth objectives when mask drift stays within ρ.

Where Pith is reading between the lines

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

  • If a stable 5-10% mask is genuinely sufficient, then the mask could be communicated to clients so they upload only masked coordinates, turning a zero-overhead accuracy fix into a communication-reduction fix; the paper does not make this claim.
  • The same projection-first recipe may transfer to federated learning outside graphs—e.g., image or text tasks with severe domain skew—because the mechanism (domain-specific variance dominating gradient coordinates) is not graph-specific; the paper tests only GNN backbones.
  • The mask is built from mean absolute gradients, so a minority domain whose informative coordinates fall below the global threshold could be systematically silenced; a fairness-aware variant that guarantees per-client coordinate coverage is a natural testable extension.
  • Given that the convergence guarantee depends on mask drift being small, a practical monitoring rule—tracking IOU between consecutive masks and pausing or widening the mask when drift spikes—could make the theoretical assumption operational.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes FedIA, a server-side aggregation method for federated graph learning under domain shift. FedIA has two stages: (1) a global top-rho mask computed from mean absolute client gradients, retaining only the most important coordinates, and (2) an influence-regularized multiplicative reweighting that down-weights clients whose masked gradients deviate from the mean. The authors report accuracy gains up to 11.58 percentage points over several FL/FGL baselines on Twitch Gamers and WikiNet, with a claimed zero extra uplink cost. They also provide a convergence sketch (Sec. 4.3, App. A.2) asserting an O(sigma^2/sqrt(T)) rate based on sparse-SGD and error-feedback results. The paper's main empirical message is that a 'projection-first' strategy — filtering noisy gradient coordinates before aggregation — can stabilize and improve FGL under domain skew.

Significance. If the empirical claims hold, FedIA is an attractive plug-in: it requires no client modification, no extra communication, and only a small server-side mask and weight vector. The reported gains across multiple baselines and two graph backbones would be practically useful. The paper also ships code, which supports reproducibility. However, the manuscript's theoretical contribution is not valid as written, and at least one central table is internally inconsistent. The significance of the paper therefore rests on the empirical results, which need to be corrected and verified.

major comments (4)
  1. [App. A.2, Eq. (10)-(11)] The claimed O(sigma^2/sqrt(T)) convergence is unsupported. The cited biased-compressor bounds (Stich et al. 2018; Karimireddy et al. 2019) require error feedback: dropped coordinates must be accumulated and re-injected. Algorithm 1 has no residual buffer, and the statement that 'our running mean \bar g_t^k' provides error feedback is a category error: in Algorithm 1, \bar g_M is the mean of masked gradients across clients, not a per-client residual. Additionally, Eq. (10) includes an additive max_t delta_t term, but Eq. (11) drops it with only the assertion delta_t <= rho, which is not proved and can be violated if the top-rho set shifts. The load-bearing theoretical claim therefore applies to a different algorithm.
  2. [Algorithm 1 vs. Sec. 4.3/App. A.2] The proof sketch assumes a server momentum update of the form alpha_k <- beta alpha_k + (1-beta) tilde_alpha_k with beta in [0.8,0.95], while Algorithm 1 (line 21) uses a multiplicative update alpha_k <- alpha_k exp(-lambda d_k) and beta never appears anywhere in the pseudocode. Table 7 nonetheless reports tuned beta values. Thus the convergence argument, even if the error-feedback issue were fixed, would not cover the implemented algorithm.
  3. [Table 4] The WikiNet table is internally inconsistent. For FedAvg+FedIA, the per-domain entries are Chameleon 25.02, Crocodile 20.55, Squirrel 21.54, whose mean is about 22.37, but the table reports AVG = 15.10 +/- 0.06. Similar discrepancies appear likely in other rows (e.g., FedProx+FedIA: (14.31+4.21+13.44)/3 = 10.65, which matches the printed 10.65, so the issue is localized but serious). The main text also says Crocodile improved to 21.54, whereas the table lists 20.55. These contradictions undermine the WikiNet empirical claim and must be corrected.
  4. [Sec. 6.1, App. B (hyperparameter selection)] The central motivating claim that 'around 5% suffices' is based on the authors' own run (Fig. 2b, 5.04%), and the operational mask ratio rho=0.1 is selected on the same two benchmarks (Table 7 and Fig. 3). Because the same data are used for tuning and for the headline comparison, the conclusion that 'a vast majority of gradient dimensions are dominated by domain-specific variance' is partly circular with respect to benchmark-specific tuning. The paper should either provide a measure of noise-dimension ratio that is independent of the final test set, or explicitly reframe this as a post-hoc explanation rather than a predictive finding.
minor comments (4)
  1. [Abstract vs. App. A.3] The abstract says FedIA 'selects important coordinates within each layer', but App. A.3 explicitly states the mask is applied globally across all parameters, not per-layer. This contradiction should be resolved.
  2. [Algorithm 1 and Table 7] Algorithm 1's Input list includes rho and lambda but not beta, although beta is a tuned hyperparameter in Table 7. The pseudocode and the hyperparameter table should agree.
  3. [Table 4, A VGΔ row] The 'A VGΔ with FedIA' row in Table 4 contains five entries ('+6.42−0.23 +3.58 +1.86 +3.40') for a table with three domains plus average, making the row unreadable.
  4. [References / formatting] There are duplicated references (Wei et al. 2021a and 2021b appear identical; Rozemberczki et al. 2021a and 2021b are the same entry) and Figure 3 contains unrecovered Unicode glyphs ('/uni00000013/...') that render as gibberish. These should be fixed.

Circularity Check

0 steps flagged

No significant circularity: FedIA's empirical benchmark claims are self-contained; the convergence sketch has an unproven mask-drift assumption but is not definitionally circular.

full rationale

The central empirical claim is self-contained: FedIA's accuracy is measured on held-out nodes against external baselines, and the mask/weighting equations (6)-(8) plus Algorithm 1 fully determine the update procedure; the reported improvements are not a logical consequence of those definitions. The '5.04% suffices' motivation and the rho=0.1 tuning are transparently empirical, presented as observations rather than as first-principles derivations. The only candidate for circularity is the convergence sketch in Appendix A.2 (Eqs. 10-11). There, the paper asserts delta_t = 1-IOU(M^{t-1},M^t) <= rho and 'confirms' it with its own Figure A1 rather than proving it, and it identifies its running mean with the error-feedback buffer required by Karimireddy et al. even though Algorithm 1 has no residual accumulator. These are unsupported steps in the theoretical justification, and the stated O(sigma^2/sqrt(T)) rate does not follow from the cited theorems as written. However, this is a correctness gap in a proof sketch, not a case where the conclusion is equivalent to the inputs by construction, nor a load-bearing self-citation chain. The paper does not import a uniqueness theorem from its own prior work, and it does not rename a known result. The empirical evaluation against external baselines remains independent support for the method. Therefore the circularity score is low, reflecting only the weak, empirically-filled convergence assumption rather than any definitional circularity in the main contribution.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on a small number of tuned parameters (rho, lambda, and the inconsistently defined beta), on the empirically motivated projection hypothesis, and on the applicability of existing sparse-SGD theory to a masked and reweighted update. No new physical or conceptual entities are introduced.

free parameters (3)
  • rho (mask ratio) = 0.1 (selected from {0.1, 0.3, 0.5, 0.7, 0.9}; varies per baseline and setting in Table 7)
    Controls how many gradient coordinates are retained. The motivating 5.04 percent comes from one FedAvg run; the deployed value is tuned post hoc.
  • lambda (influence sensitivity) = not reported, not in the search space
    Appears in Stage 2 exponential weighting in Eq. (8) and Algorithm 1, but no tuning or sensitivity analysis is provided.
  • beta (momentum factor) = Table 7 lists best beta per baseline and setting, but Algorithm 1 never uses beta
    Described as important for convergence and tuned, yet the pseudocode is inconsistent with this parameter.
axioms (4)
  • domain assumption Gradient coordinates with large mean absolute value across clients are the informative ones; top-rho projection removes domain-specific noise without losing signal.
    Core projection hypothesis, motivated by Fig. 2(b), not derived; used in Stage 1.
  • standard math Sparsified SGD convergence theorems (Stich 2018; Gao et al. 2021) apply to FedIA's masked, client-weighted update.
    The paper cites these theorems for the sparse-SGD bound in Eq. (10).
  • ad hoc to paper The mask drift delta_t = 1-IOU(M^{t-1}, M^t) is bounded by rho, so max delta_t is small enough to drop from the bound.
    Needed to obtain Eq. (11); asserted and measured in Figure A1, not proven, and it does not follow from the definition of IOU.
  • domain assumption Explicit gradient upload is available and treated as the standard client payload.
    FedIA requires clients to send raw gradients rather than weight deltas; the paper treats this as an acceptable protocol.

pith-pipeline@v1.3.0-alltime-deepseek · 26070 in / 13451 out tokens · 143237 ms · 2026-08-04T16:26:56.158413+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Dissecting Federated-Graph Aggregation under Domain Shift: Importance-Aware Aggregation via Empirical Analysis." pith.science (2026). https://pith.science/paper/MWRTCMNO

@misc{pith2026250918171,
  author       = {Pith},
  title        = {Pith review of: Dissecting Federated-Graph Aggregation under Domain Shift: Importance-Aware Aggregation via Empirical Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MWRTCMNO}},
  note         = {Machine review of arXiv:2509.18171}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Federated graph learning (FGL) trains a shared graph model across clients whose local graphs differ in node features, labels, and connectivity while keeping raw graph data decentralized. Although graph-domain shifts across clients can severely degrade the global model, existing FGL approaches for graph-domain shift mainly adapt local representations, propagation, or graph-derived collaboration, while the server typically applies the same aggregation rule to every parameter coordinate. In this work, we analyze how graph-domain shifts affect server-side aggregation. We find that, as clients optimize under distinct graph-domain conditions, they gradually concentrate their strongest updates on different parameter coordinates, making important update coordinates less shared across clients. Consequently, standard averaging can weaken these domain-specific signals even when client updates are not directly opposed. These findings reveal an aggregationinduced signal dilution effect: parameter coordinates strongly expressed by only a subset of domains are attenuated by nearzero contributions from the remaining domains. Motivated by these findings, we propose FedIA, a lightweight serverside aggregation method that calibrates parameter selection and client weighting. FedIA identifies important coordinates =within each layer and adjusts client aggregation weights according to their contributions on the selected subspace, without modifying local graph training or the standard clientupdate payload. Experiments on Twitch Gamers and WikiNet using two graph-learning backbones and nine FL/FGL baselines show improvements of up to 11.58 percentage points, with particularly strong gains under domain skew.

Figures

Figures reproduced from arXiv: 2509.18171 by Kahou Tam, Zeyu Ma, Zhanting Zhou, Ziqiang Zheng.

Figure 1
Figure 1. Figure 1: Overview of FedIA in a domain-skewed FGL scenario. (a) Domain-skewed clients (e.g., institutions from different [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Empirical evidence of core challenges in domain-skewed FGL and the effectiveness of our approach. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The effects of the hyperparameter variations are [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

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

38 extracted references · 12 linked inside Pith

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Acar, D. A. E.; Zhao, Y.; Navarro, R. M.; Mattina, M.; Whatmough, P. N.; and Saligrama, V. 2021. Federated learning based on dynamic regularization. arXiv preprint arXiv:2111.04263

    Pith/arXiv arXiv 2021

  4. [4]

    Chen, Y.; Huang, W.; and Ye, M. 2024. Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12077--12086

    2024

  5. [5]

    Cheng, Z.; Huang, X.; Wu, P.; and Yuan, K. 2023. Momentum benefits non-iid federated learning simply and provably. arXiv preprint arXiv:2306.16504

    Pith/arXiv arXiv 2023

  6. [6]

    P.; and Li, A

    Chunduru, A.; Morafah, M.; Morafah, M.; Chellapandi, V. P.; and Li, A. 2025. Avoid forgetting by preserving global knowledge gradients in federated learning with non-iid data. arXiv preprint arXiv:2505.20485

    Pith/arXiv arXiv 2025

  7. [7]

    Dai, S.; Sohn, J.-y.; Chen, Y.; Alam, S.; Balakrishnan, R.; Banerjee, S.; Himayat, N.; and Lee, K. 2024. Buffer-based Gradient Projection for Continual Federated Learning. arXiv preprint arXiv:2409.01585

    Pith/arXiv arXiv 2024

  8. [8]

    Duesing, C.; and Cimiano, P. 2025. Rethinking federated learning as a digital platform for dynamic and value-driven participation. Future Generation Computer Systems, 171: 107847

    2025

  9. [9]

    Fey, M.; and Lenssen, J. E. 2019. Fast Graph Representation Learning with PyTorch Geometric. arXiv:1903.02428

    Pith/arXiv arXiv 2019

  10. [10]

    Gao, H.; Xu, A.; and Huang, H. 2021. On the convergence of communication-efficient local SGD for federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 7510--7518

    2021

  11. [11]

    Hamilton, W.; Ying, Z.; and Leskovec, J. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems, 30

    2017

  12. [12]

    Hu, R.; Guo, Y.; and Gong, Y. 2023. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. IEEE Transactions on Mobile Computing, 23(8): 8242--8255

    2023

  13. [13]

    Huang, W.; Wan, G.; Ye, M.; and Du, B. 2024. Federated graph semantic and structural learning. arXiv preprint arXiv:2406.18937

    Pith/arXiv arXiv 2024

  14. [14]

    P.; Rebjock, Q.; Stich, S.; and Jaggi, M

    Karimireddy, S. P.; Rebjock, Q.; Stich, S.; and Jaggi, M. 2019. Error feedback fixes signsgd and other gradient compression schemes. In International conference on machine learning, 3252--3261. PMLR

    2019

  15. [15]

    Kerkouche, R.; \'A cs, G.; Castelluccia, C.; and Genev \`e s, P. 2021. Compression boosts differentially private federated learning. In 2021 IEEE European Symposium on Security and Privacy (EuroS&P), 304--318. IEEE

    2021

  16. [16]

    Lee, G.; and Choi, D. 2025. TinyProto: Communication-Efficient Federated Learning with Sparse Prototypes in Resource-Constrained Environments. arXiv preprint arXiv:2507.04327

    Pith/arXiv arXiv 2025

  17. [17]

    Li, Q.; He, B.; and Song, D. 2021. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10713--10722

    2021

  18. [18]

    K.; Zaheer, M.; Sanjabi, M.; Talwalkar, A.; and Smith, V

    Li, T.; Sahu, A. K.; Zaheer, M.; Sanjabi, M.; Talwalkar, A.; and Smith, V. 2020. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2: 429--450

    2020

  19. [19]

    McMahan, B.; Moore, E.; Ramage, D.; Hampson, S.; and y Arcas, B. A. 2017. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, 1273--1282. PMLR

    2017

  20. [20]

    Mu, X.; Shen, Y.; Cheng, K.; Geng, X.; Fu, J.; Zhang, T.; and Zhang, Z. 2023. Fedproc: Prototypical contrastive federated learning on non-iid data. Future Generation Computer Systems, 143: 93--104

    2023

  21. [21]

    Reddi, S.; Charles, Z.; Zaheer, M.; Garrett, Z.; Rush, K.; Kone c n \`y , J.; Kumar, S.; and McMahan, H. B. 2020. Adaptive federated optimization. arXiv preprint arXiv:2003.00295

    Pith/arXiv arXiv 2020

  22. [22]

    Robbins, H.; and Monro, S. 1951. A stochastic approximation method. The annals of mathematical statistics, 400--407

    1951

  23. [23]

    Rozemberczki, B.; Allen, C.; and Sarkar, R. 2021 a . Multi-scale attributed node embedding. Journal of Complex Networks, 9(2): cnab014

    2021

  24. [24]

    Rozemberczki, B.; Allen, C.; and Sarkar, R. 2021 b . Multi-scale attributed node embedding. Journal of Complex Networks, 9(2): cnab014

    2021

  25. [25]

    U.; Cordonnier, J.-B.; and Jaggi, M

    Stich, S. U.; Cordonnier, J.-B.; and Jaggi, M. 2018. Sparsified SGD with memory. Advances in neural information processing systems, 31

    2018

  26. [26]

    Sun, J.; Wu, X.; Huang, H.; and Zhang, A. 2024. On the role of server momentum in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 15164--15172

    2024

  27. [27]

    Tan, Y.; Long, G.; Jiang, J.; and Zhang, C. 2024 a . Influence-oriented personalized federated learning. arXiv preprint arXiv:2410.03315

    Pith/arXiv arXiv 2024

  28. [28]

    Tan, Y.; Long, G.; Liu, L.; Zhou, T.; Lu, Q.; Jiang, J.; and Zhang, C. 2022. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 8432--8440

    2022

  29. [29]

    Tan, Z.; Wan, G.; Huang, W.; and Ye, M. 2024 b . FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference. arXiv preprint arXiv:2410.20105

    Pith/arXiv arXiv 2024

  30. [30]

    Wan, G.; Huang, W.; and Ye, M. 2024. Federated graph learning under domain shift with generalizable prototypes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 15429--15437

    2024

  31. [31]

    Wei, W.; Liu, L.; Wu, Y.; Su, G.; and Iyengar, A. 2021 a . Gradient-leakage resilient federated learning. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), 797--807. IEEE

    2021

  32. [32]

    Wei, W.; Liu, L.; Wu, Y.; Su, G.; and Iyengar, A. 2021 b . Gradient-leakage resilient federated learning. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), 797--807. IEEE

    2021

  33. [33]

    Y.; Hu, S.; Sun, L.; and Yao, D

    Xue, L.; Hu, S.; Zhao, R.; Zhang, L. Y.; Hu, S.; Sun, L.; and Yao, D. 2024. Revisiting Gradient Pruning: A Dual Realization for Defending against Gradient Attacks. arXiv:2401.16687

    Pith/arXiv arXiv 2024

  34. [34]

    Yang, C.; Wu, Q.; Wang, J.; and Yan, J. 2022. Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps. arXiv preprint arXiv:2212.09034

    Pith/arXiv arXiv 2022

  35. [35]

    Y.; Kou, Z.; and Wang, D

    Zhang, D. Y.; Kou, Z.; and Wang, D. 2020. Fairfl: A fair federated learning approach to reducing demographic bias in privacy-sensitive classification models. In 2020 IEEE International Conference on Big Data (Big Data), 1051--1060. IEEE

    2020

  36. [36]

    ZHANG, K.; Yang, C.; Li, X.; Sun, L.; and Yiu, S. M. 2021. Subgraph Federated Learning with Missing Neighbor Generation. In Ranzato, M.; Beygelzimer, A.; Dauphin, Y.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Information Processing Systems, volume 34, 6671--6682. Curran Associates, Inc

    2021

  37. [37]

    Zhao, J.; Rao, X.; Yang, B.; Wang, Y.; He, J.; Ma, H.; Niu, W.; and Wang, W. 2025. Gradient Reconstruction Protection Based on Sparse Learning and Gradient Perturbation in IoV. International Journal of Intelligent Systems, 2025(1): 9253392

    2025

  38. [38]

    Zhu, L.; Liu, Z.; and Han, S. 2019. Deep leakage from gradients. Advances in neural information processing systems, 32

    2019