Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Low-frequency components of local SAM perturbations carry most client disagreement in federated learning; filtering them out improves accuracy and flattens the global loss.

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 · grok-4.5

2026-07-11 21:11 UTC pith:6YZHSSCN

load-bearing objection Clean empirical observation plus a zero-comm plug-in that actually moves the needle on FedSAM under hard non-IID; the spectral story is correlational but the ablations and multi-backbone results still make it worth engaging. the 3 major comments →

arxiv 2607.04170 v1 pith:6YZHSSCN submitted 2026-07-05 cs.LG

FedFFT: Taming Client Drift in Federated SAM via Spectral Perturbation Filtering

Liyang Yuan , Yibo Yang , Dandan Guo This is my paper
classification cs.LG
keywords Federated LearningSharpness-Aware MinimizationClient DriftFrequency-domain FilteringNon-IID DataSpectral Perturbation AnalysisFedFFT
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.

Federated learning fails when clients hold non-identical data because their local updates pull the shared model in conflicting directions. Sharpness-aware minimization tries to fix this by seeking flatter minima, yet the local sharpness directions themselves still disagree across clients. This paper shows, by Fourier analysis of those directions, that the disagreements concentrate in the low-frequency part of the spectrum. FedFFT simply zeros those low-frequency coefficients before the local update, keeping the higher-frequency parts that appear more consistent. The filter needs no extra communication, works as a drop-in replacement for any client-side SAM optimizer, and yields higher test accuracy and visibly flatter global landscapes, especially under severe data skew. The result matters because it turns a geometric diagnosis of client drift into a cheap, architecture-agnostic correction that scales across backbones and base algorithms.

Core claim

Inter-client variance of SAM perturbation vectors is predominantly concentrated in the low-frequency spectrum; high-pass filtering those components produces more consistent client updates, flatter global minima, and higher test accuracy than prior SAM-based federated methods under non-IID partitions.

What carries the argument

FedFFT: a per-layer real-FFT high-pass filter that zeros the lowest fraction r of frequency coefficients of each client's local SAM perturbation before the ascent step, then inverse-transforms the result.

Load-bearing premise

The premise that low-frequency coefficients of a flattened per-layer SAM perturbation mainly encode client-specific bias, while the retained high-frequency coefficients remain useful and shared learning signals.

What would settle it

On the same non-IID partitions, replace the low-frequency zeroing with an equal-sparsity high-frequency or random zeroing of the same SAM perturbations and check whether the accuracy and landscape-flatness gains disappear.

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

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

3 major / 5 minor

Summary. The paper studies client drift in federated SAM, showing via per-layer RFFT analysis that inter-client variance of local SAM perturbations concentrates in low-frequency bands under Dirichlet non-IID partitions (Fig. 1). Motivated by this, it proposes FedFFT: a plug-and-play high-pass filter (Eqs. 5–8, Algorithm 1) that zeros the lowest fraction r of frequency coefficients of each client’s SAM perturbation before the second gradient step, with no extra communication. The method is integrated with FedAvg/SCAFFOLD/FedDyn bases and evaluated on CIFAR-10/100 and Tiny-ImageNet with ResNet-18/20, DenseNet-121 and ViT under α∈{0.1,0.6}, plus ablations on filter type/ratio/radius, personalization (FedPer), aggregation (FedAWA), wall-clock efficiency and loss-landscape visualizations. Results claim consistent accuracy gains, faster convergence and flatter global minima, especially under severe heterogeneity.

Significance. If the empirical gains hold under broader scrutiny, FedFFT supplies a simple, zero-communication, architecture-agnostic module that improves existing SAM-based FL optimizers and introduces spectral analysis of optimizer signals (rather than inputs or weights) as a practical diagnostic. Strengths include the multi-dataset/multi-backbone/multi-base-algorithm experimental suite, targeted ablations (Table 5, Fig. 6), efficiency comparison (Table 3), landscape and PCA visualizations (Figs. 4–5), and demonstrated compatibility with personalization and adaptive aggregation. The absence of theory is typical for this empirical line of work; the contribution is the observation-plus-filter package and its measured effect sizes under non-IID regimes.

major comments (3)
  1. [Section 4.1 / Figure 1 / Table 5] Section 4.1 (hypothesis after Fig. 1) and the causal narrative throughout claim that low-frequency coefficients of the flattened per-layer SAM perturbation primarily carry client-specific data bias while high-frequency coefficients remain consistent learning signals. Fig. 1 only establishes higher inter-client variance in low bands; it does not distinguish client bias from shared low-frequency loss geometry whose amplitude varies across clients. Table 5 shows that low-frequency zeroing outperforms high-frequency/random/L2/clipping controls at fixed sparsity, but remains correlational: any systematic alteration of the lowest-energy coefficients could produce similar gains. Without additional diagnostics (e.g., spectral alignment with client data statistics, or comparison of clients with matched vs mismatched class distributions), the mechanistic explanation and the claim of “general appli
  2. [Table 2 / Table 4 / Figure 3] Tables 2 and 4 (and the learning curves in Fig. 3) report point estimates without standard deviations or multi-seed averages. Federated non-IID runs are known to be high-variance; several entries are author-reproduced while others are taken from prior work. Without seed-averaged results (or at least error bars on the key FedFFT vs strongest baseline comparisons), the claimed consistent outperformance—especially the large margins under α=0.1—cannot be assessed for statistical reliability.
  3. [Section 4.2, Eqs. (5)–(7)] The filter (Eqs. 5–7) flattens each weight tensor then applies a 1-D rFFT. For convolutional kernels the linearization order is arbitrary and can mix spatial frequencies; no sensitivity study to reshape order, padding, or 2-D/3-D FFT alternatives is provided. If the low-frequency concentration is an artifact of the chosen flattening, the method’s claimed spectral interpretation and transferability across architectures (including ViT) become less secure.
minor comments (5)
  1. [Algorithm 1] Algorithm 1 and the surrounding text use inconsistent indexing (δk,t,e vs. layer-wise δk_l); a short pseudocode comment clarifying that Filter is applied independently per layer would help reproducibility.
  2. [Figure 2] Figure 2 heatmaps are useful but the color-scale captions are partially garbled in the source; ensure clean legends and consistent α labels in the camera-ready version.
  3. [Section 6.1.2 / Figure 6] The recommended safe range r∈[0.5 %,4 %] (Fig. 6) is helpful; please also state the exact r used for every main-table entry so readers can reproduce without reverse-engineering.
  4. [Section 2.4] Related-work discussion of frequency analysis (Sec. 2.4) correctly cites spectral bias and FFT-gradient sparsification, but a brief sentence distinguishing “filtering for drift suppression” from “filtering for compression” would sharpen the novelty claim.
  5. [Section 4.3 / Table 3] Complexity analysis (Sec. 4.3) asserts O(d log d) is negligible; a short wall-clock breakdown of the FFT step alone (already partially in Table 3) would make the claim quantitative.

Circularity Check

0 steps flagged

No circularity: spectral observation is measured independently, filter is a free design choice, and accuracy gains are validated on held-out metrics rather than forced by construction.

full rationale

The paper's derivation chain is observational then constructive then empirical. Section 4.1 and Figure 1 report measured inter-client variance of flattened per-layer SAM perturbations (via rFFT) under Dirichlet partitions; the concentration of variance in low-frequency bands is an experimental fact, not a definition. The subsequent hypothesis (low-frequency components carry client-specific bias) motivates the high-pass Filter operator of Eqs. (5)–(8), but the operator itself is a free algorithmic choice (truncation ratio r is ablated in Fig. 6 and Table 5, not reverse-engineered from final accuracies). Downstream claims—flatter loss landscapes (Fig. 4), tighter client perturbations/features/parameters (Fig. 5), and higher test accuracy (Tables 2, 4, 6, 7)—are obtained by running the resulting optimizer and comparing against independent baselines on standard benchmarks. No equation equates the accuracy gain to the input variance statistic by construction, no parameter is fitted on the target metric and then re-presented as a prediction, and no load-bearing uniqueness or ansatz is imported via self-citation. The method is therefore self-contained against external evaluation; any remaining risk is correlational (whether low-frequency energy truly isolates client bias) rather than circular.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

The method rests on standard FL and SAM machinery plus one empirical spectral hypothesis and a small set of tunable scalars. No new physical entities are postulated; the free parameters are ordinary hyper-parameters of the proposed filter and of SAM itself.

free parameters (2)
  • frequency truncation ratio r = 0.01 (default); safe band 0.005–0.04
    Fraction of lowest-frequency coefficients zeroed; recommended range 0.5–4 % chosen by validation (Fig. 6). Directly controls the strength of the claimed intervention.
  • SAM perturbation radius ρ = 0.1 (main experiments)
    Standard SAM hyper-parameter; paper shows FedFFT is more robust to its value than plain FedSAM, yet performance still depends on it (Table 6).
axioms (3)
  • ad hoc to paper Local SAM perturbations can be treated as signals whose Fourier spectrum is meaningful after flattening each weight tensor.
    Invoked in Section 4.1–4.2; no prior theorem guarantees that the ordering of rFFT coefficients after flattening corresponds to client-specific bias.
  • domain assumption Standard FedAvg / SCAFFOLD / FedDyn aggregation and local SGD steps remain valid after the perturbation is replaced by its filtered version.
    Implicit throughout Algorithm 1 and the experimental protocol.
  • domain assumption Dirichlet-partitioned CIFAR/Tiny-ImageNet with GroupNorm ResNets are representative of realistic non-IID federated regimes.
    Standard experimental convention in the cited FL literature; used for all main claims.
invented entities (1)
  • FedFFT high-pass filter operator H_r no independent evidence
    purpose: Zeroes the lowest r-fraction of rFFT coefficients of each layer’s SAM perturbation to suppress client-specific low-frequency bias.
    Defined in Eqs. (5)–(8); the entity is an algorithmic construct whose only evidence is the empirical accuracy gains reported in the paper.

pith-pipeline@v1.1.0-grok45 · 25574 in / 2556 out tokens · 26445 ms · 2026-07-11T21:11:58.509199+00:00 · methodology

0 comments
read the original abstract

Federated Learning (FL) enables decentralized training without data sharing, but suffers from statistical heterogeneity across clients, leading to client drift, poor generalization, and sharp minima compared to centralized training. Sharpness-Aware Minimization (SAM) has emerged as a promising approach to improve generalization, yet its application in federated learning still suffers from divergence problems, since perturbations are computed locally and reflect client-specific loss geometries. To better understand this issue, we provide experimental evidence from a new perspective, the frequency domain, for SAM perturbations in federated settings, revealing that inter-client perturbation inconsistencies are predominantly concentrated in the low-frequency spectrum. Motivated by this insight, we propose Federated learning with Frequency-domain Filtering of SAM perturbations (FedFFT). It is a lightweight and plug-and-play method that filters out low-frequency components of SAM perturbations without requiring additional communication, thereby suppressing inconsistent components in client updates while preserving consistent learning signals. Extensive experiments across multiple benchmarks and diverse backbones demonstrate that FedFFT consistently outperforms SAM-based FL methods, particularly under severe non-IID distributions. These results highlight the effectiveness, scalability, and general applicability of our frequency-domain perspective for sharpness-aware federated optimization.

Figures

Figures reproduced from arXiv: 2607.04170 by Dandan Guo, Liyang Yuan, Yibo Yang.

Figure 1
Figure 1. Figure 1: (a–c) Variance of SAM perturbations across clients under different levels of data heterogeneity, controlled by the Dirichlet concentration parameter α. Across all datasets and settings, inter-client variance is consistently concentrated in the low-frequency components, indicating that client disagreement is primarily a low-frequency phe￾nomenon. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Client data distributions across CIFAR-10, CIFAR-100, and Tiny-ImageNet under Dirichlet partitions [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Learning curves on three benchmarks under Dirichlet partitionings with concentration parameters 0.1 and 0.6, using a ResNet-18 model. 5.2. Main Results Comparison with State-of-the-Art Baselines [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: 2D loss landscape comparison across different algorithms on CIFAR-10 under Dirichlet partition with α = 0.1, using ResNet-18 as the backbone model. Training Efficiency. Training efficiency is another crucial factor in federated learning (FL), as practical deployments often require methods to reach a target accuracy under limited computational and wall-clock budgets. In this study, we adopt FedAvg as the re… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of low-frequency perturbation filtering on ResNet-18/CIFAR-10 (α = 0.1). The comparison between FedSAM (blue) and FedFFT (red) shows: (a) increased compactness of client perturbations (conv1.weight), leading to (b) more consistent model features (layer4.1.gn2) and (c) more aligned client model parameters (conv1.weight). 6. Discussion 6.1. Ablation Study 6.1.1. Different filtering strategies. To vali… view at source ↗
Figure 6
Figure 6. Figure 6: Test accuracy with different filtering ratios on CIFAR-10 (α = 0.1). “-A”, “-S” and “-D” denote using FedAvg, SCAFFOLD and FedDyn as the base algorithms. 6.1.3. Robustness to FL Settings. We further conduct a systematic robustness analysis of FedFFT across three key federated learning hy￾perparameters: client activation rate, number of local training epochs, and total number of participating clients. The r… view at source ↗
Figure 7
Figure 7. Figure 7: (b) examines the effect of local computation by varying the number of local epochs from 1 to 5, with activation rate fixed at 10% and 100 clients. FedFFT exhibits the largest relative improvements at 1–2 epochs, conditions under which the optimization is more communication-constrained and convergence is harder to achieve. Although the margin decreases as local training deepens, FedFFT consistently maintain… 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

47 extracted references · 3 canonical work pages

  1. [1]

    McMahan, E

    B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, Communication-efficient learning of deep networks from decentralized data, in: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2017

    2017

  2. [2]

    R. S. Antunes, C. André da Costa, A. Küderle, I. A. Yari, B. Eskofier, Federated learning for healthcare: Systematic review and architecture proposal, ACM Trans. Intell. Syst. Technol. 13 (4) (May 2022). doi:10.1145/3501813. URLhttps://doi.org/10.1145/3501813

    doi:10.1145/3501813 2022

  3. [3]

    A.Rauniyar, D.H.Hagos, D.Jha, J.E.Håkegård, U.Bagci, D.B.Rawat, V.Vlassov, Federatedlearning for medical applications: A taxonomy, current trends, challenges, and future research directions, IEEE Internet of Things Journal 11 (5) (2024) 7374–7398.doi:10.1109/JIOT.2023.3329061

    work page doi:10.1109/jiot.2023.3329061 2024

  4. [4]

    Fantauzzo, E

    L. Fantauzzo, E. Fanì, D. Caldarola, A. Tavera, F. Cermelli, M. Ciccone, B. Caputo, Feddrive: Gen- eralizing federated learning to semantic segmentation in autonomous driving, in: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 11504–11511.doi: 10.1109/IROS47612.2022.9981098

    doi:10.1109/iros47612.2022.9981098 2022

  5. [5]

    Guan, P.-T

    H. Guan, P.-T. Yap, A. Bozoki, M. Liu, Federated learning for medical image analysis: A survey, Pattern Recognition 151 (2024) 110424.doi:10.1016/j.patcog.2024.110424. URLhttps://doi.org/10.1016/j.patcog.2024.110424

    doi:10.1016/j.patcog.2024.110424 2024

  6. [6]

    S. P. Karimireddy, S. Kale, M. Mohri, S. J. Reddi, S. U. Stich, A. T. Suresh, Scaffold: Stochastic controlled averaging for federated learning (2021).arXiv:1910.06378. URLhttps://arxiv.org/abs/1910.06378

    Pith/arXiv arXiv 2021

  7. [7]

    Woodworth, K

    B. Woodworth, K. K. Patel, S. U. Stich, Z. Dai, B. Bullins, H. B. McMahan, O. Shamir, N. Srebro, Is local sgd better than minibatch sgd?, in: Proceedings of the 37th International Conference on Machine Learning, ICML’20, JMLR.org, 2020

    2020

  8. [8]

    T. Li, A. K. Sahu, A. Talwalkar, V. Smith, Federated learning: Challenges, methods, and future directions, IEEE Signal Processing Magazine 37 (3) (2020) 50–60.doi:10.1109/MSP.2020.2975749

    doi:10.1109/msp.2020.2975749 2020

  9. [9]

    Z. Fan, Y. Wang, J. Yao, L. Lyu, Y. Zhang, Q. Tian, Fedskip: Combatting statistical heterogeneity with federated skip aggregation, in: 2022 IEEE International Conference on Data Mining (ICDM), 2022, pp. 131–140.doi:10.1109/ICDM54844.2022.00023

    doi:10.1109/icdm54844.2022.00023 2022

  10. [10]

    M. Kang, S. Kim, K. H. Jin, E. Adeli, K. M. Pohl, S. H. Park, Fednn: Federated learning on concept drift data using weight and adaptive group normalizations, Pattern Recognition 149 (2023) 110230. 17

    2023

  11. [11]

    D. A. E. Acar, Y. Zhao, R. Matas, M. Mattina, P. Whatmough, V. Saligrama, Federated learning based on dynamic regularization, in: International Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=B7v4QMR6Z9w

    2021

  12. [12]

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith, Federated optimization in heteroge- neous networks (2020).arXiv:1812.06127. URLhttps://arxiv.org/abs/1812.06127

    Pith/arXiv arXiv 2020

  13. [13]

    Q. Li, B. He, D. Song, Model-contrastive federated learning, in: 2021 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2021, pp. 10708–10717.doi:10.1109/CVPR46437. 2021.01057

    doi:10.1109/cvpr46437 2021

  14. [14]

    J. Xu, S. Wang, L. Wang, A. C.-C. Yao, Fedcm: Federated learning with client-level momentum, ArXiv abs/2106.10874 (2021). URLhttps://api.semanticscholar.org/CorpusID:235490679

    Pith/arXiv arXiv 2021

  15. [15]

    R. Ye, M. Xu, J. Wang, C. Xu, S. Chen, Y. Wang, Feddisco: federated learning with discrepancy-aware collaboration, in: Proceedings of the 40th International Conference on Machine Learning, ICML’23, JMLR.org, 2023

    2023

  16. [16]

    Z. Li, T. Lin, X. Shang, C. Wu, Revisiting weighted aggregation in federated learning with neural net- works, in: Proceedings of the 40th International Conference on Machine Learning, ICML’23, JMLR.org, 2023

    2023

  17. [17]

    C. Shi, H. Zhao, B. Zhang, M. Zhou, D. Guo, Y. Chang, Fedawa: Adaptive optimization of aggregation weights in federated learning using client vectors, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  18. [18]

    Foret, A

    P. Foret, A. Kleiner, H. Mobahi, B. Neyshabur, Sharpness-Aware Minimization for Efficiently Improving Generalization (2020).arXiv:2010.01412

    Pith/arXiv arXiv 2020

  19. [19]

    Z. Qu, X. Li, R. Duan, Y. Liu, B. Tang, Z. Lu, Generalized federated learning via sharpness aware minimization (2022).arXiv:2206.02618. URLhttps://arxiv.org/abs/2206.02618

    Pith/arXiv arXiv 2022

  20. [20]

    Caldarola, B

    D. Caldarola, B. Caputo, M. Ciccone, Improving generalization in federated learning by seeking flat minima, in: European Conference on Computer Vision, Springer, 2022, pp. 654–672

    2022

  21. [21]

    Y. Sun, L. Shen, S. Chen, L. Ding, D. Tao, Dynamic regularized sharpness aware minimization in federated learning: Approaching global consistency and smooth landscape, in: International Conference on Machine Learning, PMLR, 2023, pp. 32991–33013

    2023

  22. [22]

    Z. Fan, S. Hu, J. Yao, G. Niu, Y. Zhang, M. Sugiyama, Y. Wang, Locally estimated global perturba- tions are better than local perturbations for federated sharpness-aware minimization, in: International Conference on Machine Learning, 2024

    2024

  23. [23]

    Caldarola, P

    D. Caldarola, P. Cagnasso, B. Caputo, M. Ciccone, Beyond local sharpness: Communication-efficient global sharpness-aware minimization for federated learning, in: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2025

    2025

  24. [24]

    R. Dai, X. Yang, Y. Sun, L. Shen, X. Tian, M. Wang, Y. Zhang, Fedgamma: Federated learning with global sharpness-aware minimization, IEEE Transactions on Neural Networks and Learning Systems 35 (12) (2024) 17479–17492.doi:10.1109/TNNLS.2023.3304453

    work page doi:10.1109/tnnls.2023.3304453 2024

  25. [25]

    S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, 2011.doi:10.1561/2200000016

    doi:10.1561/2200000016 2011

  26. [26]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Simplifying neural nets by discovering flat minima, in: Advances in Neural Information Processing Systems, 1994. 18

    1994

  27. [27]

    N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, P. T. P. Tang, On large-batch training for deep learning: Generalization gap and sharp minima, in: International Conference on Learning Representa- tions, 2017

    2017

  28. [28]

    Neyshabur, S

    B. Neyshabur, S. Bhojanapalli, D. McAllester, Srebro., Exploring generalization in deep learning, in: Advances in neural information processing systems, 2017

    2017

  29. [29]

    J. Kwon, J. Kim, H. Park, I. K. Choi, Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks, in: M. Meila, T. Zhang (Eds.), Proceedings of the 38th International Conference on Machine Learning, Vol. 139 of Proceedings of Machine Learning Research, PMLR, 2021, pp. 5905–5914

    2021

  30. [30]

    Zhuang, B

    J. Zhuang, B. Gong, L. Yuan, Y. Cui, H. Adam, N. Dvornek, S. Tatikonda, J. Duncan, T. Liu, Sur- rogate gap minimization improves sharpness-aware training, in: International Conference on Learning Representations, 2022

    2022

  31. [31]

    Y. Liu, S. Mai, M. Cheng, X. Chen, C.-J. Hsieh, Y. You, Random sharpness-aware minimization., in: Advances in Neural Information Processing Systems, 2022

    2022

  32. [32]

    K.Ahn, A.Jadbabaie, S.Sra, Howtoescapesharpminimawithrandomperturbations., in: InForty-first International Conference on Machine Learning., 2024

    2024

  33. [33]

    J. Chen, H. Li, C. P. Chen, Boosting sharpness-aware training with dynamic neighborhood, Pattern Recogn. 153 (C) (Sep. 2024).doi:10.1016/j.patcog.2024.110496. URLhttps://doi.org/10.1016/j.patcog.2024.110496

    work page doi:10.1016/j.patcog.2024.110496 2024

  34. [34]

    Zhang, Z

    R. Zhang, Z. Fan, J. Yao, Y. Zhang, Y. Wang, Domain-inspired sharpness-aware minimization under domain shifts, in: International Conference on Learning Representations, 2024

    2024

  35. [35]

    Y. Sun, L. Shen, T. Huang, L. Ding, D. Tao, Fedspeed: Larger local interval, less communication round, and higher generalization accuracy (2023).arXiv:2302.10429. URLhttps://arxiv.org/abs/2302.10429

    Pith/arXiv arXiv 2023

  36. [36]

    Z. Qu, X. Li, X. Han, R. Duan, C. Shen, L. Chen, How to prevent the poor performance clients for personalized federated learning?, in: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 12167–12176.doi:10.1109/CVPR52729.2023.01171

    doi:10.1109/cvpr52729.2023.01171 2023

  37. [37]

    G. Lee, M. Jeong, S. Kim, J. Oh, S.-Y. Yun, Fedsol: Stabilized orthogonal learning with proximal restrictions in federated learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  38. [38]

    X. Xing, Q. Zhan, X. Xie, Y. Yang, Q. Wang, G. Liu, Flexible sharpness-aware personalized federated learning, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, Philadelphia, Pennsylvania, USA, 2025, pp. 21707–21715

    2025

  39. [39]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Dräxler, M. Lin, F. A. Hamprecht, Y. Bengio, A. C. Courville, On the spectral bias of neural networks, in: International Conference on Machine Learning, 2018. URLhttps://api.semanticscholar.org/CorpusID:53012119

    2018

  40. [40]

    Z.-Q. J. Xu, Y. Zhang, T. Luo, Y. Xiao, Z. Ma, Frequency principle: Fourier analysis sheds light on deep neural networks, ArXiv abs/1901.06523 (2019). URLhttps://api.semanticscholar.org/CorpusID:58981616

    Pith/arXiv arXiv 1901

  41. [41]

    D. Yin, R. G. Lopes, J. Shlens, E. D. Cubuk, J. Gilmer, A fourier perspective on model robustness in computer vision, ArXiv abs/1906.08988 (2019). URLhttps://api.semanticscholar.org/CorpusID:195317007

    Pith/arXiv arXiv 1906

  42. [42]

    H. Wang, X. Wu, P. Yin, E. P. Xing, High-frequency component helps explain the generalization of con- volutional neural networks, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 8681–8691. URLhttps://api.semanticscholar.org/CorpusID:173188317 19

    2020

  43. [43]

    L. Wang, W. Wu, Y. Zhao, J. Zhang, H. Liu, G. Bosilca, J. J. Dongarra, M. Herlihy, R. Fonseca, Superneurons: Fft-based gradient sparsification in the distributed training of deep neural networks, ArXiv abs/1811.08596 (2018). URLhttps://api.semanticscholar.org/CorpusID:53786504

    Pith/arXiv arXiv 2018

  44. [44]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton, Learning multiple layers of features from tiny images, Tech. Rep. 0, University of Toronto, Toronto, Ontario (2009). URLhttps://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf

    2009

  45. [45]

    S. U. CS231N, Tiny imagenet visual recognition challenge,http://cs231n.stanford.edu/(2015)

    2015

  46. [46]

    M. G. Arivazhagan, V. Aggarwal, A. K. Singh, S. Choudhary, Federated learning with personalization layers, CoRR abs/1912.00818 (2019).arXiv:1912.00818. URLhttp://arxiv.org/abs/1912.00818

    Pith/arXiv arXiv 1912

  47. [47]

    Caldas, S

    S. Caldas, S. M. K. Duddu, P. Wu, T. Li, J. Konečný, H. B. McMahan, V. Smith, A. Talwalkar, Leaf: A benchmark for federated settings (2019).arXiv:1812.01097. URLhttps://arxiv.org/abs/1812.01097 20

    Pith/arXiv arXiv 2019