Pith. sign in

REVIEW 4 major objections 7 minor 68 references

Robust Federated Learning against Noisy Clients via Masked Optimization

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A two-stage federated training scheme that detects noisy clients, masks suspicious samples, and corrects their labels reports stronger results than sixteen prior methods under heterogeneous label noise.

desk verdict A credible incremental extension of FedELC, but the central 'masked optimization' claim is not what the equations implement: the mask only gates entropy regularization, not the noisy-label supervision it is advertised to filter. read the letter →

arxiv 2506.02079 v1 pith:D3KNTY7O submitted 2025-06-02 cs.LG cs.AIcs.CVstat.ML

Xuefeng Jiang , Tian Wen , Zhiqin Yang , Lvhua Wu , Yufeng Chen , Sheng Sun , Yuwei Wang , Min Liu This is my paper
classification cs.LGcs.AIcs.CVstat.ML
keywords federatedlearninglabelnoisenoisyclientscorrectionmaskedoptimizationgeometricmedianaggregationGaussianmixturemodelheterogeneousdata
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

MaskedOptim tries to establish that federated learning can stay accurate when different clients hold labels corrupted at different rates, and can even improve the quality of those corrupt labels. The proposed two-stage framework first trains a warm-up model and uses per-class client losses to split clients into a relatively clean group and a relatively noisy group. Noisy clients then train with a masked, end-to-end label-correction objective: a learnable soft-label distribution is updated by backpropagation, only the 80 percent of samples with the smallest losses contribute to an entropy sharpening term, and models are aggregated with geometric median weights. Across four datasets and three real-world noise patterns, the paper reports higher precision and recall than sixteen baselines, with the clearest gains at high noise rates. A careful reader would care because label noise is common in real federated deployments, and the same pipeline both protects the model and repairs the clients' annotations.

What carries the argument

The machinery is a two-stage, two-group training loop. After $T_w$ warm-up rounds with logit-adjusted cross-entropy, each client sends per-class average losses to the server, where a two-component Gaussian mixture model assigns clients to a clean group and a noisy group. For noisy clients, every sample carries a learnable soft-label variable $\tilde{y}=K\hat{y}$ normalized by softmax, and the local objective is $L = CE(p,\tilde{y}) + \alpha\, Compatibility(\hat{y},\tilde{y}) + \beta\, Entropy(p)\cdot mask$, with the mask keeping the $\tau=80\%$ of batch samples that have the smallest supervised losses. The estimated ground-truth label is formed by averaging the model's softmax prediction with $\tilde{y}$ after the local update. Finally, the server aggregates updates by the geometric median instead of weighted averaging, so any outlier client's parameters receive smaller weight.

What would settle it

An observation that would settle the central claim: rerun the CIFAR-10 symmetric 0.0 to 0.8 protocol with a random 80 percent mask in place of the small-loss mask; if the accuracy gap over FedELC remains, the masking assumption is not what drives the result.

Watch

Extended reading notes

Core claim

The paper claims that treating noisy clients differently from clean clients is the right response to heterogeneous label noise in federated learning. Its framework, MaskedOptim, first detects noisy clients with a two-component Gaussian mixture model fitted on class-wise average losses after a warm-up stage, then optimizes the detected noisy clients with a per-sample learnable soft-label distribution masked by a small-loss filter, and finally aggregates models with the geometric median. In the decisive comparison against the closest prior method, FedELC, on CIFAR-10 with symmetric label noise from 0.0 to 0.8, MaskedOptim reaches 64.80 precision and 61.66 recall versus 60.21 and 59.77. On the large-scale Clothing1M dataset it reports a best test accuracy of 71.89 percent against 71.64 percent for FedELC. The paper also claims that its corrected labels match or exceed FedELC's label-estimation accuracy after warm-up, improving the data quality of noisy clients' local datasets.

Load-bearing premise

The framework depends on two empirical judgments: the Gaussian mixture split must put each client in the right group, and the 80 percent of samples with the smallest losses must be the correctly labeled ones; a wrong split sends clean clients through noisy-label training, and a wrong mask discards useful supervision.

Editorial extensions

If this is right

  • On CIFAR-10 with symmetric noise 0.0 to 0.8, MaskedOptim's reported precision and recall exceed FedELC's by 4.59 and 1.89 percentage points respectively, the largest gap in the synthetic-noise experiments.
  • Removing the small-loss mask or replacing geometric median aggregation with averaging, median, or distance-aware aggregation degrades the reported results, indicating both components carry the improvement.
  • On Clothing1M, where label noise is systematic and unstructured, MaskedOptim reports the best test accuracy among the compared methods at 71.89 percent.
  • The label-correction output gives noisy clients a practical re-labeling check: samples for which the estimated label differs from the original can be reviewed and re-labeled, improving local data quality.
  • On CIFAR-10-N with human annotation errors, MaskedOptim reports the highest precision and recall under IID and Dirichlet(1.0) partitioning and remains competitive at higher data heterogeneity.

Reading between the lines

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

  • The paper leaves implicit a test of how often the Gaussian mixture split puts a client in the wrong group; measuring that confusion against true per-client noise rates would show whether the reported gains come from the split itself.
  • A testable extension is an adaptive mask: because the 80 percent threshold is fixed per batch, newly corrected labels may still be masked until their losses fall, so a threshold that shrinks as correction confidence rises could improve the trade-off.
  • On the AGNews text benchmark, MaskedOptim does not beat FedProx under some noise patterns, which suggests the label-correction machinery transfers less directly to text and invites a modality-aware variant.
  • If the corrected labels are trustworthy, a deployment could freeze them as hard labels and drop the per-sample learnable variables, reducing memory and training cost.
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 / 7 minor

Summary. The paper proposes MaskedOptim, a two-stage federated learning framework for label noise. Stage one warms up a global model and then splits clients into clean and noisy groups using a two-component GMM fitted on per-class average losses. Stage two trains detected noisy clients with an end-to-end label correction mechanism: a learnable per-sample soft label distribution is initialized from the original label, optimized through a classification loss, a compatibility loss, and an entropy regularizer, with a masked entropy term and a pre-merging step. Model aggregation is performed with geometric-median weights. The method is evaluated against sixteen baselines on CIFAR-10 with synthetic noise, CIFAR-10-N with human annotation noise, AGNews with synthetic text noise, and Clothing1M with systematic noise, using precision, recall, F1, and accuracy. The paper reports improvements over the strongest prior baseline FedELC on image benchmarks, an ablation study of the mask and aggregation choices, and a label-correction accuracy comparison.

Significance. If the reported results hold, the paper provides a practical improvement over FedELC for federated learning under heterogeneous label noise, with the useful additional contribution of measuring label-correction quality on the noisy clients. The paper's strengths include evaluation on real-world noisy datasets (CIFAR-10-N, Clothing1M), an extra text-modality dataset, publicly released code, and an ablation study isolating the mask and aggregation components (Table 7). However, the central claimed mechanism of the method, as described in the Introduction and abstract, is not what the equations and Algorithm 1 implement, and several claims of superiority are not backed by variance or significance information. The contribution is therefore plausible but currently under-supported relative to the strength of the claims.

major comments (4)
  1. [Sec. 4.2, Eqs. (12)-(14); Algorithm 1] The stated mechanism of the method is that the valid mask 'filter[s] out samples with larger losses, thereby decreasing the noisy label supervision', but in the actual implementation the mask is applied only to the entropy term L_e (Eq. 13). The classification loss L_c = CE(p, y_d) (Eq. 9) and the compatibility loss L_comp (Eq. 10) are unmasked, and both carry supervision derived from the original noisy labels, directly in L_comp and through the initialization y_d = SoftMax(K * y_hat). Thus the ablation in Table 7 demonstrates that gating entropy regularization on high-loss samples helps; it does not demonstrate that the mask reduces noisy-label supervision. The central claim needs to be reworded to match the equations, or the mask needs to be applied to all loss terms that use the noisy labels, with experiments rerun accordingly.
  2. [Sec. 5.1, Tables 3 and 5] All main experiments are averaged over three seeds, but no standard deviations, confidence intervals, or significance tests are reported. Several leading differences are very small, for example CIFAR-10-N IID @Pre 87.79 vs. 87.67 for FedELC and @Rec 87.77 vs. 87.47, and CIFAR-10 symmetric 0.0-0.4 @Pre 73.44 vs. 73.03. Without dispersion measures it is not possible to verify the claim of consistent superiority over sixteen baselines. Please report per-seed results or standard deviations for the main tables and state how many random seeds each reported number represents.
  3. [Sec. 4.1, noisy client detection] The GMM-based partition into clean and noisy clients is a load-bearing component: a client placed in the wrong group receives the wrong optimization procedure, and the label-correction mechanism can in principle alter the labels of clean clients. The paper reports no detection accuracy, confusion matrix, or analysis of how detection errors affect the final accuracy. Please add an evaluation of the detection stage against the known synthetic noise rates, or otherwise provide evidence that the reported gains are robust to imperfect client separation.
  4. [Sec. 5.2, Table 4] On the AGNews text dataset, MaskedOptim is not the best method under any of the four noise patterns: FedProx achieves 65.10, 61.82, 61.15, and 63.66 F1, while MaskedOptim achieves 63.01, 57.33, 59.51, and 61.12. The paper acknowledges this in the text but the abstract and title claim robustness 'in different scenarios' without this qualification. The text-modality claim should be either qualified to image benchmarks or reframed as an exploratory finding, and the sentence in the abstract claiming general robustness should be made consistent with Table 4.
minor comments (7)
  1. [Sec. 4.2, Eq. (9) and Algorithm 1 line 20] Eq. (9) writes L_c = CE(p, y_tilde) but Algorithm 1 and Eq. (8) use y_d = SoftMax(y_tilde); please make the notation consistent between the unnormalized variable y_tilde and the normalized soft-label distribution y_d.
  2. [Sec. 4.2] The paragraph beginning 'We firstly introduce the concepts of hard/soft label' is repeated almost verbatim, which disrupts the presentation and should be merged.
  3. [Table 5] In the CIFAR-10-N Non-IID Dir(0.5) row, the entries '81.5181.35' lack a separator between the precision and recall values.
  4. [Sec. 5.1, hyperparameters] The hyperparameter selection paragraph lists alpha, beta, T_w, tau, eta, and zeta, but does not specify the value of the large constant K used in Eqs. (5)/(7) to initialize y_d; since K scales the initial label distribution, it should be reported or analyzed for sensitivity.
  5. [Sec. 3.2] The sentence 'In the prior research[], evaluations...' contains an empty citation that should be filled or removed.
  6. [Sec. 5.1, baseline settings] The text refers to 'RobustFL [47]', but RobustFL appears in the bibliography as reference [64]; [47] is RobustFed. Please correct the citation or the reference number.
  7. [Sec. 5.5, Fig. 2] The sentence 'We can also observe' at the end of the sensitivity paragraph is incomplete and should either be finished or removed.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: MaskedOptim is an incremental assembly of the authors' published FedELC/FedNoRo components plus an entropy-gating mask and RFA, and the headline accuracy claim is benchmarked against external baselines; the mask-implementation gap is a correctness concern, not circularity.

full rationale

The claimed derivation chain is not circular. The two-stage pipeline is explicitly assembled from the authors' own prior work: FedELC [14] supplies the learnable label-correction and compatibility losses, FedNoRo [55] supplies the GMM client detection, and FNBench [11] supplies the small-loss threshold tau=80; however, these are published, externally reviewed methods, so citing them is legitimate evidence rather than a reduction to the current paper's own conclusion. The central claim, that MaskedOptim outperforms sixteen baselines under heterogeneous label noise, is tested against independently implemented external methods (FedAvg, FedProx, FedExP, TrimmedMean, Krum, Median, Co-teaching, DivideMix, FedLSR, FedRN, FedCorr, ELR, etc.) on three image datasets plus AGNews, and the label-correction accuracy in Section 5.6 is measured against held-out ground-truth labels rather than derived from the method's own outputs. The most serious issue is the skeptic's implementation concern: in Eqs. 12-13 the 'valid mask' is applied only to the entropy term L_e, while L_c and L_comp, the terms that actually carry original-label supervision, remain unmasked, so the stated mechanism of filtering noisy-label supervision is not exactly what the equations implement. That is a correctness and ablation-interpretation risk, not a circularity. Score 2 reflects the heavy reliance on the authors' own FedELC as both a component and the closest baseline, but no step reduces by construction or by definition to its own inputs.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

MaskedOptim is an empirical method paper: it introduces seven hand-set hyperparameters (alpha, beta, tau, eta, zeta, T_w, K, with K never specified) and relies on four domain assumptions (GMM client separation, small-loss mask validity, convergence of the learnable label distribution, geometric-median robustness). The per-sample learnable label distribution is a trainable variable, not an independently evidenced entity. The strongest assumptions are inherited from the authors' own prior work FedELC and FedNoRo.

free parameters (7)
  • alpha (trade-off coefficient for compatibility loss) = 0.5
    Fixed by hand in Section 5.1; sensitivity shown in Fig. 2 but not discussed in text.
  • beta (trade-off coefficient for entropy loss) = 0.1
    Fixed by hand in Section 5.1; sensitivity shown in Fig. 2 but not discussed in text.
  • tau (mask filter rate) = 80%
    Eq. 12; described as 'mildly fixed' with reference to benchmark [11], not tuned per dataset.
  • eta (learning rate for label distribution) = 1000
    Eq. 15; fixed at 1000 in Section 5.1 with no derivation or sensitivity analysis.
  • zeta (pre-merging rate) = 0.8
    Eq. 16; fixed in Section 5.1; sensitivity shown in Fig. 2 but not discussed in text.
  • T_w (warm-up rounds) = 30
    Section 5.1; sensitivity shown in Fig. 2 but not discussed in text.
  • K (scaling constant for label distribution) = unspecified
    Eq. 5 and Algorithm 1 line 30 call K 'a large constant'; no value is ever reported, yet it scales the learnable logits and affects the label-correction update.
assumptions (5)
  • domain assumption Small-loss samples are more likely to be correctly labeled
    Used to build the valid mask in Eq. 12, citing Co-teaching [8]; no verification is provided for the federated noisy-client setting.
  • domain assumption A two-component GMM on per-class average losses separates clean from noisy clients
    Section 4.1 adopts this from FedNoRo and FedELC; detection accuracy is never measured or reported.
  • domain assumption The learnable distribution y~ initialized from the noisy label converges toward the true label under the three-term loss
    Section 4.2 and Eq. 15; label-correction accuracy is only reported as a figure (Fig. 3) without numerical values for the final corrected labels.
  • domain assumption Geometric median aggregation reduces the influence of outlier client models in this FL setting
    Algorithm 2 is taken from RFA [42]; the ablation in Table 7 supports it empirically on one setting only.
  • standard math Logit adjustment with the local class prior improves training under class imbalance
    Eq. 4, citing [38]; used in warm-up and clean-client updates; established externally.
invented entities (1)
  • Per-sample learnable label distribution y~ (soft label)
    purpose: Provides a trainable estimate of each sample's ground-truth label on detected noisy clients, updated via backpropagation (Eq. 5-8 and Eq. 15).
    The variable is introduced and justified only within the paper; its correctness is demonstrated indirectly through final accuracy and a label-correction figure (Fig. 3) with no numerical values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Federated Learning against Noisy Clients via Masked Optimization." pith.science (2026). https://pith.science/paper/D3KNTY7O

@misc{pith2026250602079,
  author       = {Pith},
  title        = {Pith review of: Robust Federated Learning against Noisy Clients via Masked Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3KNTY7O}},
  note         = {Machine review of arXiv:2506.02079}
}
read the original abstract

In recent years, federated learning (FL) has made significant advance in privacy-sensitive applications. However, it can be hard to ensure that FL participants provide well-annotated data for training. The corresponding annotations from different clients often contain complex label noise at varying levels. This label noise issue has a substantial impact on the performance of the trained models, and clients with greater noise levels can be largely attributed for this degradation. To this end, it is necessary to develop an effective optimization strategy to alleviate the adverse effects of these noisy clients.In this study, we present a two-stage optimization framework, MaskedOptim, to address this intricate label noise problem. The first stage is designed to facilitate the detection of noisy clients with higher label noise rates. The second stage focuses on rectifying the labels of the noisy clients' data through an end-to-end label correction mechanism, aiming to mitigate the negative impacts caused by misinformation within datasets. This is achieved by learning the potential ground-truth labels of the noisy clients' datasets via backpropagation. To further enhance the training robustness, we apply the geometric median based model aggregation instead of the commonly-used vanilla averaged model aggregation. We implement sixteen related methods and conduct evaluations on three image datasets and one text dataset with diverse label noise patterns for a comprehensive comparison. Extensive experimental results indicate that our proposed framework shows its robustness in different scenarios. Additionally, our label correction framework effectively enhances the data quality of the detected noisy clients' local datasets. % Our codes will be open-sourced to facilitate related research communities. Our codes are available via https://github.com/Sprinter1999/MaskedOptim .

Figures

Figures reproduced from arXiv: 2506.02079 by the authors.

Figure 1
Figure 1. Illustration of MaskedOptim. 4.1 Noisy Client Detection Our method is divided into two stages where the first stage lasts for 𝑇𝑤 global communication rounds to warm up the global model. During this stage, a warm-up model is firstly trained based on FedAvg [37]. When the warm-up training ends, we exploit a two-component Gaussian Mixture Model (GMM) to divide total clients into a relatively clean group and a relativel… view at source ↗
Figure 2
Figure 2. Visualization of the sensitivity study. 5.6 Analysis on Label Correction We reckon considering noisy labels in FL, we can not only improve the trained model’s robustness against label noise but also try to conduct label correction. In [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the label correction performance. negative impacts of the noisy labels, we construct a valid mask to filter out the possible misinformation from imperfect label supervision to further enhance the robustness of the trained model. To further mitigate the weight divergence in the model aggregation phase, we exploit the geometric median based model aggregation mechanism to decrease adverse impacts cause… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 41 canonical work pages

  1. [1]

    Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel

    David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. 2019. MixMatch: A Holistic Approach to Semi-Supervised Learning. InAdvances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, Hanna M....

    work page 2019

  2. [2]

    Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. 2017. Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent. InAdvances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, Isabelle Guyon, Ulrike von Luxburg, Samy Be...

    work page 2017

  3. [3]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition. Ieee, 248–255

    2009

  4. [4]

    2016.Regulation (EU) 2016/679 of the European Parliament and of the Council

    European Parliament and Council of the European Union. 2016.Regulation (EU) 2016/679 of the European Parliament and of the Council. https: //data.europa.eu/eli/reg/2016/679/oj

    work page 2016

  5. [6]

    Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. 2017. Robust loss functions under label noise for deep neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 31

    2017

  6. [7]

    Yves Grandvalet and Yoshua Bengio. 2004. Semi-supervised Learning by Entropy Minimization. InAdvances in Neural Information Processing Systems 17 [Neural Information Processing Systems, NIPS 2004, December 13-18, 2004, Vancouver, British Columbia, Canada]. 529–536. https: //proceedings.neurips.cc/paper/2004/hash/96f2b50b5d3613adf9c27049b2a888c7-Abstract.html

    work page 2004

  7. [8]

    Tsang, and Masashi Sugiyama

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. InAdvances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Can...

    work page 2018

  8. [9]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. IEEE Computer Society, 770–778. doi:10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

Show all 68 references
  1. [10]

    Divyansh Jhunjhunwala, Shiqiang Wang, and Gauri Joshi. 2023. FedExP: Speeding Up Federated Averaging via Extrapolation. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net. https://openreview.net/pdf?id=...

    2023

  2. [11]

    Xuefeng Jiang, Jia Li, Nannan Wu, Zhiyuan Wu, Xujing Li, Sheng Sun, Gang Xu, Yuwei Wang, Qi Li, and Min Liu. 2024. FNBench: Benchmarking Robust Federated Learning against Noisy Labels.Authorea Preprints(2024)

    2024

  3. [12]

    Xuefeng Jiang, Peng Li, Sheng Sun, Jia Li, Lvhua Wu, Yuwei Wang, Xiuhua Lu, Xu Ma, and Min Liu. 2025. Refining Distributed Noisy Clients: An End-to-end Dual Optimization Framework.Authorea Preprints(2025)

    2025

  4. [13]

    Xuefeng Jiang, Yuan Ma, Pengxiang Li, Leimeng Xu, Xin Wen, Kun Zhan, Zhongpu Xia, Peng Jia, XianPeng Lang, and Sheng Sun. 2025. TransDiffuser: End-to-end Trajectory Generation with Decorrelated Multi-modal Representation for Autonomous Driving.arXiv preprint arXiv:2505.09315(2025)

    2025

  5. [14]

    Xuefeng Jiang, Sheng Sun, Jia Li, Jingjing Xue, Runhan Li, Zhiyuan Wu, Gang Xu, Yuwei Wang, and Min Liu. 2024. Tackling Noisy Clients in Federated Learning with End-to-end Label Correction. InProceedings of the 33rd ACM International Conference on Information and Knowledge Man...

    2024

  6. [16]

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of Tricks for Efficient Text Classification. InProceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers. doi:10.18653/v1/e17-2068

    2017 doi

  7. [17]

    Hui Kang, Sheng Liu, Huaxi Huang, Jun Yu, Bo Han, Dadong Wang, and Tongliang Liu. 2023. Unleashing the Potential of Regularization Strategies in Learning with Noisy Labels.CoRRabs/2307.05025 (2023). doi:10.48550/arXiv.2307.05025 arXiv:2307.05025

    2023 doi arXiv

  8. [18]

    Sangmook Kim, Wonyoung Shin, Soohyuk Jang, Hwanjun Song, and Se-Young Yun. 2022. FedRN: Exploiting k-Reliable Neighbors Towards Robust Federated Learning. InProceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17...

    2022

  9. [19]

    2024.FLR: Label-Mixture Regularization for Federated Learning with Noisy Labels

    Taehyeon Kim, Donggyu Kim, and Se-Young Yun. 2024.FLR: Label-Mixture Regularization for Federated Learning with Noisy Labels. https: //openreview.net/pdf?id=Z8A3HDgS0E

    2024

  10. [20]

    Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images. (2009)

    2009

  11. [21]

    Jia Li, Lijie Hu, Zhixian He, Jingfeng Zhang, Tianhang Zheng, and Di Wang. 2024. Text Guided Image Editing with Automatic Concept Locating and Forgetting.CoRRabs/2405.19708 (2024). doi:10.48550/ARXIV.2405.19708 arXiv:2405.19708 Manuscript submitted to ACM Robust Federated Lear...

    2024 doi arXiv

  12. [22]

    Jia Li, Lijie Hu, Jingfeng Zhang, Tianhang Zheng, Hua Zhang, and Di Wang. 2025. Fair Text-to-Image Diffusion via Fair Mapping. InAAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA, Toby Walsh...

    2025 doi

  13. [23]

    Junyi Li, Jian Pei, and Heng Huang. 2022. Communication-Efficient Robust Federated Learning with Noisy Labels. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022, Aidong Zhang and Huzefa Rangwala (Eds.). ...

    2022

  14. [24]

    Junnan Li, Richard Socher, and Steven C. H. Hoi. 2020. DivideMix: Learning with Noisy Labels as Semi-supervised Learning. In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. https://openreview.net/for...

    2020

  15. [25]

    Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. 2022. Federated Learning on Non-IID Data Silos: An Experimental Study. In38th IEEE International Conference on Data Engineering, ICDE 2022, Kuala Lumpur, Malaysia, May 9-12, 2022. IEEE, 965–978. doi:10.1109/ICDE53745.2022.00077

    2022

  16. [26]

    Qinbin Li, Bingsheng He, and Dawn Song. 2021. Model-Contrastive Federated Learning. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. Computer Vision Foundation / IEEE, 10713–10722. doi:10.1109/CVPR46437.2021.01057

    2021

  17. [27]

    Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. 2021. Ditto: Fair and Robust Federated Learning Through Personalization. InProceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event (Proceedings of Machine Learning...

    2021

  18. [28]

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. 2020. Federated Optimization in Heterogeneous Networks. InProceedings of Machine Learning and Systems 2020, MLSys 2020, Austin, TX, USA, March 2-4, 2020, Inderjit S. Dhillon, Dimitris...

    2020

  19. [30]

    Xujing Li, Sheng Sun, Min Liu, Ju Ren, Xuefeng Jiang, and Tianliu He. 2024. FedCRAC: Improving Federated Classification Performance on Long-Tailed Data via Classifier Representation Adjustment and Calibration.IEEE Transactions on Mobile Computing(2024)

    2024

  20. [31]

    Siqi Liang, Jintao Huang, Dun Zeng, Junyuan Hong, Jiayu Zhou, and Zenglin Xu. 2023. FedNoisy: Federated Noisy Label Learning Benchmark. CoRRabs/2306.11650 (2023). doi:10.48550/arXiv.2306.11650 arXiv:2306.11650

    2023 doi arXiv

  21. [32]

    Ziqian Lin, Xuefeng Jiang, Kun Zhang, Chongjun Fan, and Yaya Liu. 2025. FedDSHAR: A dual-strategy federated learning approach for human activity recognition amid noise label user.Future Gener. Comput. Syst.166 (2025), 107724. doi:10.1016/J.FUTURE.2025.107724

    2025

  22. [33]

    Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. 2020. Early-Learning Regularization Prevents Memorization of Noisy Labels. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, Neu...

    2020

  23. [34]

    Xiuhua Lu, Peng Li, and Xuefeng Jiang. 2025. FedLF: Adaptive Logit Adjustment and Feature Optimization in Federated Long-Tailed Learning. In Asian Conference on Machine Learning. PMLR, 303–318

    2025

  24. [35]

    Yang Lu, Lin Chen, Yonggang Zhang, Yiliang Zhang, Bo Han, Yiu-ming Cheung, and Hanzi Wang. 2023. Federated Learning with Extremely Noisy Clients via Negative Distillation.arXiv preprint arXiv:2312.12703(2023)

    2023 arXiv

  25. [36]

    McLachlan and Suren I

    Geoffrey J. McLachlan and Suren I. Rathnayake. 2014. On the number of components in a Gaussian mixture model.WIREs Data Mining Knowl. Discov.4, 5 (2014), 341–355. doi:10.1002/WIDM.1135

    2014 doi

  26. [37]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20...

    2017

  27. [38]

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. 2021. Long-tail learning via logit adjustment. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.ne...

    2021

  28. [39]

    Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. 2018. Mixed Precision Training. InInternational Conference on Learning Representations

    2018

  29. [40]

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch. (2017)

    2017

  30. [41]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. 2011. Scikit-learn: Machine learning in Python.the Journal of machine Learning research12 (2011), 2825–2830

    2011

  31. [42]

    Krishna Pillutla, Sham M Kakade, and Zaid Harchaoui. 2022. Robust aggregation for federated learning.IEEE Transactions on Signal Processing70 (2022), 1142–1154

    2022

  32. [43]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  33. [44]

    Ramaswamy, SingYu Lin, Dora Zhao, AaronB

    VikramV. Ramaswamy, SingYu Lin, Dora Zhao, AaronB. Adcock, Laurensvander Maaten, Deepti Ghadiyaram, and Olga Russakovsky. 2023. GeoDE: a Geographically Diverse Evaluation Dataset for Object Recognition. (Jan 2023)

    2023

  34. [45]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  35. [46]

    Hwanjun Song, Minseok Kim, and Jae-Gil Lee. 2019. SELFIE: Refurbishing Unclean Samples for Robust Deep Learning. InProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA (Proceedings of Machine Learning Rese...

    2019

  36. [47]

    Farnaz Tahmasebian, Jian Lou, and Li Xiong. 2022. RobustFed: A Truth Inference Approach for Robust Federated Learning. InProceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, Mohammad Al Hasan and Li ...

    2022

  37. [48]

    Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. 2018. Joint Optimization Framework for Learning With Noisy Labels. In2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. Computer Vision Found...

    2018

  38. [49]

    Vasileios Tsouvalas, Aaqib Saeed, Tanir Ozcelebi, and Nirvana Meratnia. 2023. Labeling Chaos to Learning Harmony: Federated Learning with Noisy Labels.ACM Transactions on Intelligent Systems and Technology(2023)

    2023

  39. [50]

    Yuwei Wang, Runhan Li, Hao Tan, Xuefeng Jiang, Sheng Sun, Min Liu, Bo Gao, and Zhiyuan Wu. 2023. Federated skewed label learning with logits fusion.arXiv preprint arXiv:2311.08202(2023)

    2023 arXiv

  40. [51]

    Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. 2019. Symmetric Cross Entropy for Robust Learning With Noisy Labels. In2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019. IEEE, 3...

    2019

  41. [52]

    Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. 2022. Learning with Noisy Labels Revisited: A Study Using Real-World Human Annotations. InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Op...

    2022

  42. [53]

    Tian Wen, Hanqing Zhang, Han Zhang, Huixin Wu, Danxin Wang, Xiuwen Liu, Weishan Zhang, Yuwei Wang, and Shaohua Cao. 2024. RTIFed: A Reputation based Triple-step Incentive mechanism for energy-aware Federated learning over battery-constricted devices.Computer Networks241 (2024), 110192

    2024

  43. [54]

    Nannan Wu, Zhaobin Sun, Zengqiang Yan, and Li Yu. 2024. FedA3I: annotation quality-aware aggregation for federated medical image segmentation against heterogeneous annotation noise. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 15943–15951

    2024

  44. [55]

    Nannan Wu, Li Yu, Xuefeng Jiang, Kwang-Ting Cheng, and Zengqiang Yan. 2023. Fednoro: Towards noise-robust federated learning by addressing class imbalance and label noise heterogeneity.arXiv preprint arXiv:2305.05230(2023)

    2023 arXiv

  45. [56]

    Yangyang Xiang, Nannan Wu, Li Yu, Xin Yang, Kwang-Ting Cheng, and Zengqiang Yan. 2024. FedIA: Federated Medical Image Segmentation with Heterogeneous Annotation Completeness. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 373–382

    2024

  46. [57]

    Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. 2015. Learning from massive noisy labeled data for image classification. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015. IEEE Computer Society, 2691–2699. ...

    2015

  47. [58]

    Jingyi Xu, Zihan Chen, Tony Q. S. Quek, and Kai Fong Ernest Chong. 2022. FedCorr: Multi-Stage Federated Learning for Label Noise Correction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. IEEE, 10174–10183. ...

    2022

  48. [59]

    Jingjing Xue, Min Liu, Sheng Sun, Yuwei Wang, Hui Jiang, and Xuefeng Jiang. 2023. FedBIAD: Communication-Efficient and Accuracy-Guaranteed Federated Learning with Bayesian Inference-Based Adaptive Dropout. InIEEE International Parallel and Distributed Processing Symposium, IPD...

    2023

  49. [60]

    Jingjing Xue, Sheng Sun, Min Liu, Qi Li, and Ke Xu. 2025. Enhancing Federated Learning Robustness using Locally Benignity-Assessable Bayesian Dropout.IEEE Transactions on Information Forensics and Security(2025)

    2025

  50. [61]

    Bingjie Yan, Danmin Cao, Xinlong Jiang, Yiqiang Chen, Weiwei Dai, Fan Dong, Wuliang Huang, Teng Zhang, Chenlong Gao, Qian Chen, et al. 2024. FedEYE: A scalable and flexible end-to-end federated learning platform for ophthalmology.Patterns5, 2 (2024)

    2024

  51. [62]

    Bingjie Yan, Qian Chen, Yiqiang Chen, Xinlong Jiang, Wuliang Huang, Bingyu Wang, Zhirui Wang, Chenlong Gao, and Teng Zhang. 2024. Buffalo: Biomedical Vision-Language Understanding with Cross-Modal Prototype and Federated Foundation Model Collaboration. InProceedings of the 33r...

    2024

  52. [64]

    Seunghan Yang, Hyoungseob Park, Junyoung Byun, and Changick Kim. 2022. Robust Federated Learning With Noisy Labels.IEEE Intell. Syst.37, 2 (2022), 35–43. doi:10.1109/MIS.2022.3151466

    2022

  53. [65]

    Rui Ye, Yaxin Du, Zhenyang Ni, Siheng Chen, and Yanfeng Wang. 2023. Fake it till make it: Federated learning with consensus-oriented generation. arXiv preprint arXiv:2312.05966(2023). Manuscript submitted to ACM Robust Federated Learning against Noisy Clients via Masked Optimi...

    2023 arXiv

  54. [66]

    Kun Yi and Jianxin Wu. 2019. Probabilistic end-to-end noise correction for learning with noisy labels. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7017–7025

    2019

  55. [67]

    Bartlett

    Dong Yin, Yudong Chen, Kannan Ramchandran, and Peter L. Bartlett. 2018. Byzantine-Robust Distributed Learning: Towards Optimal Statistical Rates. InProceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15,...

    2018

  56. [68]

    Tsang, and Masashi Sugiyama

    Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor W. Tsang, and Masashi Sugiyama. 2019. How does Disagreement Help Generalization against Label Corruption?. InProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, Californi...

    2019

  57. [69]

    Jingwen Zhang, Yuezhou Wu, and Rong Pan. 2021. Incentive Mechanism for Horizontal Federated Learning Based on Reputation and Reverse Auction. InWWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021, Jure Leskovec, Marko Grobelnik, Marc Najork...

    2021

  58. [70]

    Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015. Character-level Convolutional Networks for Text Classification. InAdvances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, C...

    2015

  59. [71]

    Yonggang Zhang, Zhiqin Yang, Xinmei Tian, Nannan Wang, Tongliang Liu, and Bo Han. 2024. Robust Training of Federated Models with Extremely Label Deficiency. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=qxLVaYbsSI

    2024

  60. [72]

    Zhilu Zhang and Mert R. Sabuncu. 2018. Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels. InAdvances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018,...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.