Pith. sign in

REVIEW 3 major objections 5 minor 57 references

A Post-Processing-Based Fair Federated Learning Framework

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Standard FL training plus client-local debiasing gives large equalized-odds reductions with little accuracy cost.

desk verdict A simple, useful empirical framework for local fairness post-processing in FL; the abstract oversells the accuracy cost, but the core contribution is real and worth reviewing. read the letter →

arxiv 2501.15318 v1 pith:FAAEJLEX submitted 2025-01-25 cs.LG cs.AIcs.CY

Yi Zhou , Naman Goel This is my paper · ORCID
classification cs.LGcs.AIcs.CY
keywords federatedlearningalgorithmicfairnessequalizedoddsgrouppost-processingdataheterogeneityFedAvgfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that group fairness in federated learning does not have to be built into the training objective. Instead, clients can first train a global model with a standard unconstrained FL algorithm such as FedAvg, and then each client separately debias its own copy of the model on its local data. The authors demonstrate two local debiasing tools: output post-processing, which probabilistically flips predictions to satisfy equalized odds, and final-layer fine-tuning, which adds a fairness loss while freezing all other layers. Across tabular, ECG, and chest X-ray datasets, they report large reductions in equalized-odds gaps with small accuracy cost, and in some heterogeneous settings fine-tuning improves accuracy while reducing bias. The payoff is that fairness becomes a per-client, context-dependent decision with no extra communication rounds and no sensitive-attribute sharing during global training.

What carries the argument

The load-bearing mechanism is the temporal and spatial decoupling of fairness from training: a global model is learned without any fairness constraint, and only afterward does each client apply a local mapping from that shared model to its own fair predictor. Two concrete instantiations carry the experiments: the output post-processor, which solves a linear program to find four probabilities $p_{ya} = \Pr(\tilde Y = 1 \mid \hat Y = y, A = a)$ that define a randomized derived predictor satisfying equalized odds while minimizing expected loss; and final-layer fine-tuning, which freezes all layers except the last and optimizes $\alpha l + l'$, where $l$ is an accuracy loss and $l'$ penalizes the difference in true-positive and false-positive rates between sensitive groups. The first mechanism changes only predictions, the second only the last-layer weights, and neither requires additional communication rounds or sensitive statistics at the server.

What would settle it

Run the framework on a partition with $\alpha=0.1$, or any split in which at least one client has an empty cell in the label-by-sensitive-attribute table, and measure client-level equalized odds and accuracy: if the fairness gains vanish or accuracy collapses on those clients, the claim that the method is especially effective in more heterogeneous settings fails.

Watch

Extended reading notes

Core claim

The central claim is that a standard FedAvg model, trained with no fairness constraints, can be turned into a fair model per client by a completely local post-processing stage. Each client computes its own derived predictor by solving the equalized-odds linear program from [17] on its local data, or fine-tunes only the final layer with a combined accuracy-plus-fairness loss. Because the fairness intervention happens after and separately from global training, clients are free to choose different fairness metrics, thresholds, sensitive attributes, or no intervention at all. The paper reports that this two-stage recipe reduces weighted-average equalized odds by roughly 79-87% on COMPAS and PTB-XL with output post-processing, and that final-layer fine-tuning can improve balanced accuracy while reducing bias, especially under high data heterogeneity. The authors also report that the framework's benefits shrink when data partitions are near-i.i.d. or when the unconstrained model is already fairly unbiased, as on the NIH Chest X-Ray dataset.

Load-bearing premise

Each client must have enough labeled examples in every combination of sensitive group and true label for its local debiasing step to estimate the needed rates; the paper itself notes that the strongest heterogeneity setting ($\alpha=0.1$) had to be dropped because some clients had only one or zero samples in a group.

Editorial extensions

If this is right

  • Output post-processing after FedAvg reduces weighted-average equalized odds by about 79-87% on COMPAS and PTB-XL while adding almost no training time and no extra communication rounds.
  • Final-layer fine-tuning can improve fairness with minimal accuracy loss, and under high heterogeneity ($\alpha=0.5$) it can raise both fairness and balanced accuracy at the same time.
  • Clients can enforce different fairness definitions, thresholds, or sensitive attributes independently, or skip fairness entirely, because the global training stage never sees fairness information.
  • The framework inherits the communication cost of plain FedAvg and avoids the slower, more communication-heavy fairness-aware aggregation used by FairFed-style baselines.
  • The gains are largest when the initial model is biased and local partitions are heterogeneous; on near-i.i.d. splits or datasets where FedAvg already has low EOD, the improvement is smaller.

Reading between the lines

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

  • Inference: because the local debiasing step needs enough samples in each (label, sensitive-attribute) cell, the headline claim of being especially effective in heterogeneous settings should be read as applying only down to the point where every client retains all four cells; below that, the method cannot even measure the fairness it targets.
  • Inference: the two-stage design suggests a natural deployment policy where a client estimates local EOD first and only spends its post-processing budget when the gap exceeds a context-dependent threshold, avoiding pointless accuracy loss on already-fair clients.
  • Inference: the same decoupling could be applied to other group fairness metrics, such as demographic parity, or to non-binary labels, by swapping the local post-processor, since the global model carries no fairness commitment.
  • Inference: in production, per-client post-processing means the same global model can produce different decisions for the same individual depending on which client serves them; that is a feature for local autonomy but a governance question the paper does not address.
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

3 major / 5 minor

Summary. The paper proposes a two-stage framework for fair federated learning: first, a global model is trained with standard FedAvg without fairness constraints; second, each client independently applies a local post-processing debiasing method on its own data. Two instantiations are studied: model output post-processing (PP) following Hardt et al., and final-layer fine-tuning (FT) with a fairness-aware loss. The framework is evaluated on Adult, COMPAS, PTB-XL ECG, and NIH Chest X-Ray datasets under Dirichlet heterogeneity levels alpha=0.5, 5, 500, against FedAvg, FairFed, and FairFed/FR baselines, reporting client-level and weighted-average EOD, accuracy or balanced accuracy, training time, and communication rounds. The authors claim that the framework simplifies fairness implementation in FL and provides significant fairness improvements with minimal accuracy loss or even accuracy gain across data modalities, especially under high heterogeneity.

Significance. If the empirical claims are appropriately calibrated, the paper offers a genuinely simple and decentralized alternative to fair-FL methods that impose global fairness constraints: no fairness information is exchanged during training, no additional communication rounds are needed, and clients retain the flexibility to choose their own fairness definitions and levels. A notable strength is that the post-processors are fitted on each client's local training set but all reported EOD and accuracy numbers are on held-out local test sets, so the central evaluation is not circular. The paper also provides code, covers tabular, signal, and image modalities, and includes timing and communication comparisons. The main weaknesses are that the headline 'minimal accuracy loss' claim is contradicted by the paper's own PTB-XL and COMPAS tables, the 'especially effective in more heterogeneous settings' claim is not consistently supported across methods and datasets, and variance is reported only in Figure 3, not in the main per-client tables.

major comments (3)
  1. [Abstract; Section 7.4; Section 7.6; Table 10] The abstract's claim of 'significant fairness improvements with minimal accuracy loss or even accuracy gain, across data modalities and machine learning methods' is not supported by the reported results. In Table 10, PP on PTB-XL reduces weighted EOD from 0.342/0.342/0.345 to 0.044/0.055/0.046, but weighted balanced accuracy falls from 0.790/0.770/0.770 to 0.684/0.648/0.645 for alpha=0.5/5/500, i.e. relative decreases of about 13%, 16%, and 16%. Section 7.6 itself states that 'PP comes with drop in accuracy in nearly all cases,' and Table 7 shows COMPAS accuracy dropping from 0.670 to 0.610 at alpha=0.5. The only instantiation with accuracy gains is FT on PTB-XL, but there the EOD reductions are only 12%, 7%, and 8%. The abstract therefore overgeneralizes a method- and dataset-dependent trade-off; the claims should be reworded to state which method delivers which combination of benefits, or the experiments should be extended to identify regimes where PP preserves accuracy.
  2. [Section 7.1; Section 7.6; Tables 7 and 11] The statement that the framework is 'especially effective in more heterogeneous settings' is not consistently supported. On COMPAS (Table 7), PP's relative EOD improvement over FedAvg is about 79% at alpha=0.5 but about 85% at alpha=500, so PP is slightly less effective at the highest heterogeneity level. On NIH Chest X-Ray (Table 11), PP's relative EOD improvement is about 36% at alpha=0.5, about 12% at alpha=5, and about 27% at alpha=500, which is not monotonic in heterogeneity. Only FT shows a consistent trend, with relative improvements of 39%, 18%, and 9% across alpha=0.5, 5, 500. Section 7.5's sentence that fairness improvement 'decreases with the decrease of data heterogeneity' is therefore inaccurate for PP. The heterogeneity claim should be qualified by method and dataset, or the analysis should explain why PP behaves non-monotonically.
  3. [Section 7.2; Tables 7, 8, 10, 11] The main empirical tables report only averages over 10 random seeds, with no standard deviations or confidence intervals, even though Figure 3 demonstrates that standard deviations are substantial for FairFed and non-negligible for other methods. Since the paper's central claim is that the framework 'consistently' improves fairness and that baselines are less stable, the absence of variance measures in Tables 7, 8, 10, and 11 makes it impossible to assess whether the reported EOD improvements and accuracy differences are statistically meaningful. At minimum, weighted-average EOD and accuracy/BA should be reported with standard deviations or confidence intervals for every method and heterogeneity level.
minor comments (5)
  1. [Section 7.1; Figure 3 caption] The text says 'The first row of Figure 3 shows the EOD values' and 'The second row of Figure 3 provides the test accuracy,' but the caption states that row 1 is test accuracy and row 2 is EOD; the text and caption should be reconciled.
  2. [Section 7.3; Table 8] The text mentions that 'FT method is also efficient with less than half of the training time of FairFed method,' yet Table 8 reports no FT results for the Adult dataset; either FT results should be included or the sentence should be removed.
  3. [Table 10] The communication-round values for FairFed are inconsistent across heterogeneity levels: 164 rounds at alpha=0.5 but 324 rounds at alpha=5 and alpha=500, with no explanation of why the FairFed baseline uses a different number of rounds at the highest heterogeneity level.
  4. [Section 2] There is a typo in the FairFed description: 'clients send send their local updates' should read 'clients send their local updates.'
  5. [Section 3] The phrase 'different post-process debiasing methods methods' contains a duplicated word and should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported fairness and accuracy numbers are measured on held-out local test sets after the local post-processor is fit, so the headline improvements are not forced by construction.

full rationale

The paper's only 'fit-to-data' components are the per-client derived predictor p_k (Eq. 4), which solves an LP minimizing expected loss subject to equalized-odds constraints on the local training set, and the last-layer fine-tuning objective in Algorithm 2. If the paper reported training-set EOD, those improvements would be close to forced by construction. However, every headline table (Tables 7, 8, 10, 11) reports client-level test accuracy/balanced accuracy and EOD on held-out local test sets, which are split 80/20 as described in Section 5.2, so the claimed fairness improvements and accuracy changes are empirical rather than definitional. The framework does not rename a fitted parameter as a prediction; the derived predictor is a standard post-processor applied at inference time. The self-citation in reference [2] is cited only as background on fairness-aware FL training and is not load-bearing for the framework's claims. The internal tension between the abstract's 'minimal accuracy loss' and the paper's own PTB-XL PP results is a correctness or overclaiming issue, not circularity. Therefore no circular step can be exhibited.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The framework introduces no new entities and no fitted constants in its definition; it reuses Hardt et al. post-processing and last-layer fine-tuning. The empirical claims rest on hand-selected hyperparameters (alpha_ft, beta) and on local data quality assumptions, with one important free parameter (fine-tuning rounds R) left unspecified.

free parameters (3)
  • alpha_ft (fine-tuning fairness weight) = Adult 1.0, COMPAS 2.0, PTB-XL 1.0, NIH-Chest 0.1
    Chosen per dataset by hand (Table 6); controls the fairness-accuracy trade-off in Algorithm 2. No sensitivity analysis is reported.
  • beta (FairFed fairness budget) = COMPAS 0.5, others 0.1
    Baseline hyperparameter selected to give the best reported FairFed results; affects the strength of the comparison.
  • Number of local fine-tuning rounds R = Not specified
    Algorithm 2 runs R rounds, but no value or selection procedure is given for any dataset, so the FT results are not fully reproducible from the paper alone.
assumptions (3)
  • standard math Hardt et al. equalized-odds post-processing solves the stated LP optimally on the local training set.
    Invoked in Section 4.1 when each client computes p_k = EqOdds(Yhat_k, D_k).
  • domain assumption Each client's local dataset contains the sensitive attribute and labels, so the post-processor can be fit locally.
    Algorithm 1 line 11 and Algorithm 2 line 12 require (Y, A) from D_k; the framework offers no path when these are absent.
  • domain assumption The local 80/20 train/test split is representative of each client's deployment distribution, so a post-processor fit on train generalizes to test.
    Section 5.2 defines the split; the reported EOD and accuracy are on the local test sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Post-Processing-Based Fair Federated Learning Framework." pith.science (2026). https://pith.science/paper/FAAEJLEX

@misc{pith2026250115318,
  author       = {Pith},
  title        = {Pith review of: A Post-Processing-Based Fair Federated Learning Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FAAEJLEX}},
  note         = {Machine review of arXiv:2501.15318}
}
read the original abstract

Federated Learning (FL) allows collaborative model training among distributed parties without pooling local datasets at a central server. However, the distributed nature of FL poses challenges in training fair federated learning models. The existing techniques are often limited in offering fairness flexibility to clients and performance. We formally define and empirically analyze a simple and intuitive post-processing-based framework to improve group fairness in FL systems. This framework can be divided into two stages: a standard FL training stage followed by a completely decentralized local debiasing stage. In the first stage, a global model is trained without fairness constraints using a standard federated learning algorithm (e.g. FedAvg). In the second stage, each client applies fairness post-processing on the global model using their respective local dataset. This allows for customized fairness improvements based on clients' desired and context-guided fairness requirements. We demonstrate two well-established post-processing techniques in this framework: model output post-processing and final layer fine-tuning. We evaluate the framework against three common baselines on four different datasets, including tabular, signal, and image data, each with varying levels of data heterogeneity across clients. Our work shows that this framework not only simplifies fairness implementation in FL but also provides significant fairness improvements with minimal accuracy loss or even accuracy gain, across data modalities and machine learning methods, being especially effective in more heterogeneous settings.

Figures

Figures reproduced from arXiv: 2501.15318 by the authors.

Figure 1
Figure 1. A sample 12-lead ECG record from PTB-XL dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. Experiment using COMPAS dataset with different [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 26 canonical work pages

  1. [12]

    Yuying Duan, Yijun Tian, Nitesh Chawla, and Michael Lemmon. 2024. Post-Fair Federated Learning: Achieving Group and Community Fairness in Federated Learning via Post-processing. arXiv preprint arXiv:2405.17782 (2024)

    arXiv 2024

  2. [1]

    Python Imaging Library

    2024. Python Imaging Library. https://en.wikipedia.org/w/index.php?title= Python_Imaging_Library&oldid=1225854061 Page Version ID: 1225854061

    work page 2024

  3. [2]

    Annie Abay, Yi Zhou, Nathalie Baracaldo, Shashank Rajamoni, Ebube Chuba, and Heiko Ludwig. 2020. Mitigating Bias in Federated Learning. https://doi. org/10.48550/arXiv.2012.02447 arXiv:2012.02447 [cs, stat]

    arXiv doi:10.48550/arxiv.2012.02447 2020

  4. [3]

    Durmus Alp Emre Acar, Yue Zhao, Ruizhao Zhu, Ramon Matas, Matthew Mat- tina, Paul Whatmough, and Venkatesh Saligrama. 2021. Debiasing Model Up- dates for Improving Personalized Federated Training. In Proceedings of the 38th International Conference on Machine Learning . PMLR, 21–31. https: //proceedings.mlr.press/v139/acar21a.html ISSN: 2640-3498

    work page 2021

  5. [4]

    Abien Fred Agarap. 2019. Deep Learning using Rectified Linear Units (ReLU). https://doi.org/10.48550/arXiv.1803.08375 arXiv:1803.08375 [cs, stat]

    arXiv doi:10.48550/arxiv.1803.08375 2019

  6. [5]

    Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023. Fairness and machine learning: Limitations and opportunities . MIT press

    2023

  7. [6]

    Ronny Kohavi Barry Becker. 1996. Adult. https://doi.org/10.24432/C5XW20

    doi:10.24432/c5xw20 1996

  8. [7]

    Rachel K. E. Bellamy, Kuntal Dey, Michael Hind, Samuel C. Hoffman, Stephanie Houde, Kalapriya Kannan, Pranay Lohia, Jacquelyn Martino, Sameep Mehta, Aleksandra Mojsilovic, Seema Nagar, Karthikeyan Natesan Ramamurthy, John Richards, Diptikalyan Saha, Prasanna Sattigeri, Moninder Singh, Kush R. Varsh- ney, and Yunfeng Zhang. 2018. AI Fairness 360: An Extens...

    arXiv doi:10.48550/arxiv.1810.01943 2018

Show all 57 references
  1. [8]

    Reuben Binns. 2020. On the apparent conflict between individual and group fairness. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAT* ’20). Association for Computing Machinery, New York, NY, USA, 514–524. https://doi.org/10.1145/3351095.3372864

    2020

  2. [9]

    Kay Henning Brodersen, Cheng Soon Ong, Klaas Enno Stephan, and Joachim M. Buhmann. 2010. The Balanced Accuracy and Its Posterior Distribution. In 2010 20th International Conference on Pattern Recognition . 3121–3124. https://doi.org/ 10.1109/ICPR.2010.764 ISSN: 1051-4651. 13

    2010 doi

  3. [10]

    Xin Che, Jingdi Hu, Zirui Zhou, Yong Zhang, and Lingyang Chu. 2024. Training Fair Models in Federated Learning without Data Privacy Infringement. In 2024 IEEE International Conference on Big Data (BigData) . IEEE, 7687–7696

    2024

  4. [11]

    Wei Du, Depeng Xu, Xintao Wu, and Hanghang Tong. 2020. Fairness-aware Agnostic Federated Learning. https://doi.org/10.48550/arXiv.2010.05057 arXiv:2010.05057 [cs]

    2020 doi arXiv

  5. [13]

    Sanghamitra Dutta, Dennis Wei, Hazar Yueksel, Pin-Yu Chen, Sijia Liu, and Kush Varshney. 2020. Is there a trade-off between fairness and accuracy? a perspective using mismatched hypothesis testing. In International conference on machine learning. PMLR, 2803–2813

    2020

  6. [14]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Rich Zemel

  7. [15]

    Ezzeldin, Shen Yan, Chaoyang He, Emilio Ferrara, and A

    Yahya H. Ezzeldin, Shen Yan, Chaoyang He, Emilio Ferrara, and A. Salman Avestimehr. 2023. FairFed: Enabling Group Fairness in Federated Learning. Proceedings of the AAAI Conference on Artificial Intelligence 37, 6 (June 2023), 7494–7502. https://doi.org/10.1609/aaai.v37i6.2591...

    2023 doi

  8. [16]

    Thomas Grote and Geoff Keeling. 2022. Enabling fairness in healthcare through machine learning. Ethics and Information Technology 24, 3 (2022), 39

    2022

  9. [17]

    Moritz Hardt, Eric Price, and Nathan Srebro. 2016. Equality of Opportunity in Su- pervised Learning. https://doi.org/10.48550/arXiv.1610.02413 arXiv:1610.02413 [cs]

    2016 doi arXiv

  10. [18]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Identity Map- pings in Deep Residual Networks. https://doi.org/10.48550/arXiv.1603.05027 arXiv:1603.05027 [cs]

    2016 doi arXiv

  11. [19]

    Shen, George Shih, Ronald M

    Gregory Holste, Song Wang, Ziyu Jiang, Thomas C. Shen, George Shih, Ronald M. Summers, Yifan Peng, and Zhangyang Wang. 2022. Long-Tailed Classification of Thorax Diseases on Chest X-Ray: A New Benchmark Study. In Data Augmen- tation, Labelling, and Imperfections , Hien V. Nguy...

    2022

  12. [20]

    Kenneth Holstein, Jennifer Wortman Vaughan, Hal Daumé III, Miro Dudik, and Hanna Wallach. 2019. Improving fairness in machine learning systems: What do industry practitioners need?. In Proceedings of the 2019 CHI conference on human factors in computing systems . 1–16

    2019

  13. [21]

    Sergey Ioffe and Christian Szegedy. 2015. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. https://doi.org/ 10.48550/arXiv.1502.03167 arXiv:1502.03167 [cs]

    2015 doi arXiv

  14. [22]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Opti- mization. https://doi.org/10.48550/arXiv.1412.6980 arXiv:1412.6980 [cs]

    2017 doi arXiv

  15. [23]

    Agostina J Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. 2020. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences 117, 23 (2020), 12592–12594

    2020

  16. [24]

    Li Li, Yuxi Fan, Mike Tse, and Kuo-Yi Lin. 2020. A review of applications in federated learning. Computers & Industrial Engineering 149 (2020), 106854

    2020

  17. [25]

    Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. 2021. Ditto: Fair and Robust Federated Learning Through Personalization. In Proceedings of the 38th International Conference on Machine Learning . PMLR, 6357–6368. https: //proceedings.mlr.press/v139/li21h.html ISSN: 2640-3498

    2021

  18. [26]

    Tian Li, Maziar Sanjabi, Ahmad Beirami, and Virginia Smith. 2020. Fair Resource Allocation in Federated Learning. https://doi.org/10.48550/arXiv.1905.10497 arXiv:1905.10497 [cs, stat]

    2020 doi arXiv

  19. [27]

    Lima, Antônio H

    Emilly M. Lima, Antônio H. Ribeiro, Gabriela M. M. Paixão, Manoel Horta Ribeiro, Marcelo M. Pinto-Filho, Paulo R. Gomes, Derick M. Oliveira, Ester C. Sabino, Bruce B. Duncan, Luana Giatti, Sandhi M. Barreto, Wagner Meira Jr, Thomas B. Schön, and Antonio Luiz P. Ribeiro. 2021. ...

    2021 doi

  20. [28]

    Caio B. S. Maior, João M. M. Santana, Isis D. Lins, and Márcio J. C. Moura. 2021. Convolutional neural network model based on radiological images to support COVID-19 diagnosis: Evaluating database biases. PLOS ONE 16, 3 (March 2021), e0247839. https://doi.org/10.1371/journal.p...

    2021 doi

  21. [29]

    Disha Makhija, Xing Han, Joydeep Ghosh, and Yejin Kim. 2024. Achieving fairness across local and global models in federated learning. arXiv preprint arXiv:2406.17102 (2024)

    2024 arXiv

  22. [30]

    Anqi Mao, Mehryar Mohri, and Yutao Zhong. 2023. Cross-Entropy Loss Func- tions: Theoretical Analysis and Applications. https://doi.org/10.48550/arXiv. 2304.07288 arXiv:2304.07288 [cs, stat]

    2023 doi arXiv

  23. [31]

    Yuzhen Mao, Zhun Deng, Huaxiu Yao, Ting Ye, Kenji Kawaguchi, and James Zou. 2023. Last-Layer Fairness Fine-tuning is Simple and Effective for Neural Networks. https://doi.org/10.48550/arXiv.2304.03935 arXiv:2304.03935 [cs]

    2023 doi arXiv

  24. [32]

    Ricards Marcinkevics, Ece Ozkan, and Julia E. Vogt. 2022. Debiasing Deep Chest X-Ray Classifiers using Intra- and Post-processing Methods. In Proceedings of the 7th Machine Learning for Healthcare Conference . PMLR, 504–536. https: //proceedings.mlr.press/v182/marcinkevics22a....

    2022

  25. [33]

    Mattu, Jeff Larson, Julia Angwin, Lauren Kirchner, and Surya. 2016. How We Analyzed the COMPAS Recidivism Algorithm. ProPublica (2016). https://www. propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm

    2016

  26. [34]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics. PMLR, 1273–1282...

    2017

  27. [35]

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. 2022. A Survey on Bias and Fairness in Machine Learning. Comput. Surveys 54, 6 (July 2022), 1–35. https://doi.org/10.1145/3457607

    2022 doi

  28. [36]

    Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. 2022. Local Learning Matters: Rethinking Data Heterogeneity in Federated Learning. 8397–8406. https://openaccess.thecvf. com/content/CVPR2022/html/Mendieta_Local_Learning_Matters_Rethinking_ ...

    2022

  29. [37]

    Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. 2019. Agnostic Federated Learning. InProceedings of the 36th International Conference on Machine Learning. PMLR, 4615–4625. https://proceedings.mlr.press/v97/mohri19a.html ISSN: 2640-3498

    2019

  30. [38]

    Peter A Noseworthy, Zachi I Attia, LaPrincess C Brewer, Sharonne N Hayes, Xiaoxi Yao, Suraj Kapa, Paul A Friedman, and Francisco Lopez-Jimenez. 2020. Assessing and mitigating bias in medical artificial intelligence: the effects of race and ethnicity on a deep learning model fo...

    2020

  31. [39]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019 doi arXiv

  32. [40]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard Duchesna...

    2011

  33. [41]

    Mana Saleh Al Reshan, Kanwarpartap Singh Gill, Vatsala Anand, Sheifali Gupta, Hani Alshahrani, Adel Sulaiman, and Asadullah Shaikh. 2023. Detection of Pneumonia from Chest X-ray Images Utilizing MobileNet Model. Healthcare 11, 11 (Jan. 2023), 1561. https://doi.org/10.3390/heal...

    2023 doi

  34. [42]

    Sebastian Ruder. 2017. An overview of gradient descent optimization algorithms. https://doi.org/10.48550/arXiv.1609.04747 arXiv:1609.04747 [cs]

    2017 doi arXiv

  35. [43]

    Teresa Salazar, Helder Araújo, Alberto Cano, and Pedro Henriques Abreu. 2024. A Survey on Group Fairness in Federated Learning: Challenges, Taxonomy of Solutions and Directions for Future Research. arXiv preprint arXiv:2410.03855 (2024)

    2024 arXiv

  36. [44]

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang- Chieh Chen. 2019. MobileNetV2: Inverted Residuals and Linear Bottlenecks. https://doi.org/10.48550/arXiv.1801.04381 arXiv:1801.04381 [cs]

    2019 doi arXiv

  37. [45]

    Aviv Shamsian, Aviv Navon, Ethan Fetaya, and Gal Chechik. 2021. Personalized Federated Learning using Hypernetworks. InProceedings of the 38th International Conference on Machine Learning . PMLR, 9489–9502. https://proceedings.mlr. press/v139/shamsian21a.html ISSN: 2640-3498

    2021

  38. [46]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research 15, 56 (2014), 1929–1958. http://jmlr.org/papers/v15/srivastava14a.html

    2014

  39. [47]

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Wojciech Samek, and Tobias Schaeffter. 2022. PTB-XL, a large publicly available electrocardiography dataset. https://doi.org/10.13026/KFZX-AW45

    2022 doi

  40. [48]

    Tongnian Wang, Kai Zhang, Jiannan Cai, Yanmin Gong, Kim-Kwang Raymond Choo, and Yuanxiong Guo. 2024. Analyzing the Impact of Personalization on Fairness in Federated Learning for Healthcare. Journal of Healthcare Informatics Research 8, 2 (June 2024), 181–205. https://doi.org/...

    2024 doi

  41. [49]

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. 2017. ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. In 2017 IEEE Conference on Computer ...

    2017 arXiv

  42. [50]

    Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. 2019. Bayesian Nonparametric Federated Learning of Neural Networks. https://doi.org/10.48550/arXiv.1905. 12022 arXiv:1905.12022 [cs, stat]

    2019 doi arXiv

  43. [51]

    Yuzi He, Keith Burghardt, and Kristina Lerman. 2020. A Geometric Solution to Fair Representations. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (2020). https://dl.acm.org/doi/abs/10.1145/3375627.3375864

    2020

  44. [52]

    Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P. Gummadi. 2017. Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification without Disparate Mistreatment. In Proceedings of the 26th International Conference on World Wide Web . 11...

    2017

  45. [53]

    Yuchen Zeng, Hongxu Chen, and Kangwook Lee. 2022. Improving Fairness via Federated Learning. https://doi.org/10.48550/arXiv.2110.15545 arXiv:2110.15545 [cs]

    2022 doi arXiv

  46. [54]

    Chen Zhang, Yu Xie, Hang Bai, Bin Yu, Weihong Li, and Yuan Gao. 2021. A survey on federated learning. Knowledge-Based Systems 216 (2021), 106775

    2021

  47. [55]

    Daniel Yue Zhang, Ziyi Kou, and Dong Wang. 2020. FairFL: A Fair Federated Learning Approach to Reducing Demographic Bias in Privacy-Sensitive Clas- sification Models. In 2020 IEEE International Conference on Big Data (Big Data) . 1051–1060. https://doi.org/10.1109/BigData50022...

    2020

  48. [56]

    Joshua C Zhao, Saurabh Bagchi, Salman Avestimehr, Kevin S Chan, Somali Chaterji, Dimitris Dimitriadis, Jiacheng Li, Ninghui Li, Arash Nourian, and Holger R Roth. 2024. Federated Learning Privacy: Attacks, Defenses, Applications, and Policy Landscape-A Survey. arXiv preprint ar...

    2024 arXiv

  49. [2011]

    https://doi.org/10.48550/arXiv.1104.3913 arXiv:1104.3913 [cs]

    Fairness Through Awareness. https://doi.org/10.48550/arXiv.1104.3913 arXiv:1104.3913 [cs]

    doi arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.