acceptodds
Under review as a conference paper at ICLR 2027

Globally coordinated sample reweighting for fair federated learning under heterogeneity

Abstract

Federated learning (FL) enables collaborative model training across decentralized institutions, making group fairness an important requirement when the resulting global model is deployed across demographically heterogeneous users. However, assessing global group fairness requires statistics aggregated across clients, whereas FL updates the model using private local data. This inherent mismatch makes it challenging to guarantee global group fairness, especially under client heterogeneity. To address this problem, we introduce GloReFair, a globally coordinated sample reweighting framework that bridges this gap by converting global fairness feedback into local training signals. The key idea is to separate where fairness correction is needed from how training emphasis should be allocated. We realize this separation through a server-side controller that adjusts channel prices using global fairness residuals. Given these prices, clients compute sample weights using a unique closed-form rule derived from the Lagrangian of an idealized round-wise allocation problem. We further establish theoretical guarantees on centralized equivalence and strict controller expressivity. Across three federated benchmarks, GloReFair achieves favorable fairness-utility tradeoffs and remains robust as client heterogeneity becomes severe, including settings in which competing fair-FL methods substantially degrade.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated learning, group fairness, adaptive sample reweighting, heterogeneity

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_S6CAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GloReFair, a globally coordinated sample-reweighting framework for FL, which aims to address global fairness disparities by translating global fairness objectives into sample-level attention signals during local training. By coordinating sample reweighting across clients, GloReFair seeks to achieve fairness correction through decentralized optimization.

Critical strengths
  1. The paper addresses fairness in federated learning, which is an important and practically relevant research problem. The proposed approach of translating global fairness objectives into sample-level reweighting during local training provides a reasonable and potentially effective mechanism for fairness-aware federated optimization.
  2. The paper provides theoretical analysis to support the effectiveness of the proposed method under specified assumptions and constraints.
Critical weaknesses
  1. The theoretical analysis primarily focuses on binary classification tasks, leaving the effectiveness and generalizability of GloReFair to more complex learning scenarios insufficiently established. In particular, under limited local data availability and highly non-IID distributions, estimating group-specific rates in multi-class classification tasks may suffer from substantial statistical variance. The authors do not adequately demonstrate whether the proposed fairness correction mechanism remains reliable under such challenging conditions.
  2. The residual signals are computed using the arithmetic mean of group-specific rates, which may overlook inherent differences in prediction difficulty across tasks or classes. Additionally, assigning equal importance to different groups without considering their respective sample sizes may introduce estimation bias or amplify the influence of statistically unreliable groups.
  3. The proposed sample-reweighting mechanism appears to require sample-level attention computation during local training, potentially introducing substantial computational overhead. In particular, if the attention weights must be calculated individually for each training sample in every communication round, the additional computation could significantly increase local training costs and reduce the overall efficiency of federated optimization.
Questions for authors
  1. The theoretical analysis primarily considers binary classification tasks. How does the proposed framework generalize to multi-class classification problems, particularly under highly non-IID data distributions and limited local sample sizes?
  2. The residual signals are calculated using the arithmetic mean of group-specific rates. What is the theoretical or empirical justification for assigning equal importance to different groups or tasks, particularly when they exhibit different levels of prediction difficulty and substantially different sample sizes?
  3. How does GloReFair ensure reliable fairness estimation when certain sensitive groups or classes contain very few samples on individual clients?
  4. Does the proposed attention-based reweighting mechanism require computing an individual attention weight for every training sample in each communication round? If so, what is the additional computational complexity compared with standard local training? Could the authors provide a detailed analysis of the per-round computational overhead, including additional training time and memory consumption?
6 comments
AuthorsOpenReview ↗

[W1, Q1, Q3] Concerns on binary classification and limited local data

Thank you for your comments. Our binary classification setting follows the established definitions of equal opportunity and equalized odds, providing a precise setting for studying global group fairness under federated heterogeneity. A concrete route to multiclass learning is a classwise one-versus-rest formulation: each binary subtask would have its own group-label channels, channel prices, and pooled fairness residuals. This extends the same principle of translating global fairness feedback into local training signals.

GloReFair directly addresses sparse local group-label coverage through pooled statistics. Clients contribute sample counts, positive-prediction counts, and score sums, from which the server computes global rates. A locally absent channel contributes zero statistics and requires no local rate estimate. Supplement B establishes that, for a fixed pooled dataset, common model, prediction rule, and mixture coefficient under full participation, the fairness feedback is invariant to client partitioning. Consequently, dividing the same data into smaller local subsets leaves this feedback unchanged. Its statistical support is determined by the pooled sample count in each channel.

The experiments provide complementary evidence at the training level. Supplement H.4 examines increasing the number of clients from 5 to 25 while keeping the total training set fixed, directly testing greater local data fragmentation. The following table presents the original EOD results under the strongest evaluated joint Dirichlet heterogeneity, where both demographic composition and label distributions vary substantially across clients.

Table: EOD results under joint Dirichlet heterogeneity with H = 0.2.

A dash denotes failed runs.

MethodAdult Accuracy ↑Adult EOD gap ↓CreditCard Accuracy ↑CreditCard EOD gap ↓Dutch Accuracy ↑Dutch EOD gap ↓
FedAvg0.804 ± 0.0280.349 ± 0.2340.792 ± 0.0090.144 ± 0.1260.806 ± 0.0200.476 ± 0.087
FairFed0.790 ± 0.0330.471 ± 0.1240.792 ± 0.0080.138 ± 0.1110.797 ± 0.0220.517 ± 0.173
FedFB0.806 ± 0.0260.288 ± 0.1400.792 ± 0.0080.134 ± 0.1000.733 ± 0.0420.021 ± 0.018
BGL0.797 ± 0.0070.139 ± 0.0350.735 ± 0.0730.088 ± 0.0940.824 ± 0.0060.078 ± 0.037
Function Tracking0.812 ± 0.0250.102 ± 0.0800.790 ± 0.0070.098 ± 0.0640.831 ± 0.0020.167 ± 0.074
LoGoFair0.342 ± 0.1690.616 ± 0.1200.727 ± 0.1380.399 ± 0.3090.429 ± 0.2180.523 ± 0.147
FedFACT-in------------
FedFACT-post------------
GloReFair0.820 ± 0.0080.092 ± 0.0500.789 ± 0.0050.039 ± 0.0090.824 ± 0.0020.060 ± 0.005

These results demonstrate a stable fairness-accuracy trade-off under severe client heterogeneity, consistent with the partition-invariant construction of its global fairness feedback.

AuthorsOpenReview ↗

[W2, Q2] Distinction between group weighting and statistical reliability

Thank you for raising the distinction between group weighting and statistical reliability. We use the arithmetic mean because it provides a symmetric reference for comparing the conditional prediction rates of different groups. As defined in Section~3, TPRs are averaged across groups within the positive-label slice, and FPRs are averaged separately within the negative-label slice. Subtracting this reference produces centered residuals that preserve the differences between group rates and sum to zero within each slice. All residuals vanish exactly when the corresponding group rates coincide, directly matching the EO/EOD parity objective. The construction compares groups under the same label condition, without requiring positive and negative classes to have the same prediction difficulty.

Equal importance reflects the equal standing of sensitive groups under the adopted fairness criteria. EO/EOD requires the corresponding conditional rates to agree across groups, regardless of their prevalence in the dataset. A smaller group has the same claim to fair treatment as a larger group. We therefore give each active group equal weight in defining the fairness reference. Group size determines how much evidence is available to estimate a rate; it does not determine the importance of that group's fairness. Differences in prediction difficulty can affect the attainable fairness--accuracy trade-off, while the parity target remains the same.

Sampling uncertainty is a separate statistical issue. When a group-label channel has few observations in the pooled dataset, its empirical rate can be an imprecise estimate of the population rate. This limitation is shared by finite-sample estimation of group fairness, including centralized EO/EOD evaluation. Its existence does not by itself establish estimation bias caused by equal-group weighting. GloReFair uses sample counts explicitly: each channel rate is computed from prediction statistics and counts aggregated across all clients. For a fixed common model and prediction rule under full participation, Supplement~B establishes that these rates equal those computed centrally on the pooled data. Consequently, client partitioning introduces no additional error into the empirical rates; the remaining sampling uncertainty depends on the global support available in each channel.

AuthorsOpenReview ↗

[W3, Q4] Computational overhead of attention computation

Thank you for raising the concerns on computational overhead of GloReFair. GloReFair computes each sample's weight once per reweighting round and reuses it throughout all local SGD steps. Clients evaluate the sample losses at the synchronized model and apply the closed-form weighting rule element-wise, with gradients disabled during weight preparation. For a fixed model, weight construction and storage scale linearly with the local sample count. Each stored weight occupies one FP32 value.

We measured the computational cost on Adult with LR classifier, five fully participating clients, joint Dirichlet heterogeneity H = 0.2, and the EOD objective. The three methods shared data partitions, initial models, and batch schedules. The following table reports the aggregate computation time of the five-client simulation, including weight preparation, local optimization, aggregation, fairness feedback, and controller updates.

Table: Per-round computation time (seconds) across local training budgets.

Entries are means ± sample standard deviations over three seeds.

Local epochsFedAvgFedFBGloReFair
10.079 ± 0.0010.100 ± 0.0030.144 ± 0.0014
50.393 ± 0.0030.446 ± 0.0020.599 ± 0.004
100.769 ± 0.0100.872 ± 0.0111.180 ± 0.010
201.532 ± 0.0031.719 ± 0.0182.319 ± 0.014

The results identify where this computation is spent. At 10 local epochs, weight preparation, including its forward evaluation, takes 10.900 milliseconds per round and accounts for 0.92% of GloReFair's total algorithm time. Most of the additional time occurs in local optimization, which includes weighted-loss evaluation, and the proximal penalty. The measured overall overhead is 53.4% relative to FedAvg and 35.4% relative to FedFB. Thus, the principal computational cost lies in the local optimization stage, while the sample weights are constructed once and reused across the round.

Table: Cost breakdown at 10 local epochs.

Stage times are means over three seeds in milliseconds per round. Peak memory is reported as mean ± sample standard deviation in MiB and covers the entire process.

ComponentFedAvgFedFBGloReFair
Weight preparation (ms)0.0000.00010.900
Local optimization (ms)768.781859.9301152.877
Aggregation and distribution (ms)0.5200.5200.525
Fairness feedback (ms)0.00010.91415.227
Server control (ms)0.0000.1470.705
Peak process memory (MiB)569.3 ± 0.1570.1 ± 0.5573.0 ± 0.5
Reviewer_nupuOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GloReFair, a globally coordinated sample reweighting framework for improving group fairness in heterogeneous federated learning. The server aggregates group–label statistics to construct fairness residuals and updates shared channel prices, while clients combine these prices with individual sample losses to determine local training weights. The paper derives the weighting rule from an idealized allocation problem and analyzes partition invariance, single-step centralized equivalence, controller expressivity, and controller boundedness and regret. Experiments on Adult, Default Credit Card, and Dutch under multiple heterogeneity settings demonstrate favorable fairness–accuracy tradeoffs, with particularly encouraging robustness under strong demographic heterogeneity.

Critical strengths
  • The paper addresses an important challenge in fair federated learning: fairness is assessed over the pooled population, whereas optimization occurs on heterogeneous local datasets.

  • Substantial empirical evidence of robustness under the evaluated heterogeneity settings. The experiments cover three benchmarks, both EO and EOD, and group, label, and joint Dirichlet partitions across several heterogeneity levels. The comparison includes relevant methods with different fairness mechanisms.

  • An interpretable formulation supported by complementary analysis and ablations. The allocation formulation gives a clear interpretation of the sample-weighting rule, while the theoretical results explain the roles of pooled feedback, shared prices, and separate label-channel control.

Critical weaknesses
  • Limited fairness and convergence guarantees for the implemented algorithm. The paper does not establish a direct fairness guarantee for the final classifier. The conditional bound in Appendix G concerns the smooth residuals of a score ensemble under pure smooth feedback and relies on an additional model-comparison assumption that is not established for the proposed updates. The controller’s boundedness and sublinear regret do not establish convergence of the coupled model–controller training procedure, particularly with multiple local updates. Consequently, whether the implemented algorithm converges to a solution satisfying the intended EO/EOD constraints remains theoretically unresolved.
  • Limited experimental breadth. All experiments use logistic regression on three tabular binary-classification datasets (Adult, CreditCard, and Dutch), with sex as the sole sensitive attribute and full client participation. Although the evaluation covers several heterogeneity settings and includes a client-count study from 5 to 25 clients, it does not examine neural network models, other data modalities, or partial-participation settings. Consequently, the experiments provide limited evidence that the benefits of the proposed reweighting mechanism extend beyond the evaluated model class and federated setting.
Questions for authors
  1. Under what conditions can the proposed training procedure guarantee the final classifier’s hard EO/EOD performance?

  2. Can the authors establish convergence of the coupled model-controller updates when clients perform multiple local optimization steps? What assumptions are needed to ensure that the interaction between local training, parameter aggregation, and price adaptation leads to a solution satisfying the intended fairness constraints?

  3. Do the reported fairness-accuracy gains persist with neural network classifiers and datasets beyond tabular binary classification? Additional evidence in these settings would help establish whether the proposed reweighting mechanism remains effective across different model classes and data modalities.

  4. How does GloReFair perform when only a subset of clients participates in each round, particularly when participation is correlated with demographic composition? Can the controller still obtain representative global fairness feedback, and how does incomplete feedback affect fairness and training stability?

6 comments
AuthorsOpenReview ↗

[W1] Distinction between theoretic guarantee and implemented algorithm

Thank you for identifying this distinction. We agree that the current analysis does not establish convergence of the implemented model–controller procedure to a final classifier satisfying hard EO/EOD constraints, particularly with multiple local SGD steps. We have made this limitation explicit in the Conclusion and Appendix G.

Our theoretical results characterize specific components of GloReFair. The centralized-equivalence result applies to one synchronized full-batch step. The controller result establishes boundedness and sublinear regret against fixed feasible controllers on the realized feedback sequence. These properties explain aspects of the mechanism, but do not establish convergence of the coupled training dynamics.

Regarding Appendix G, we agree with the reviewer’s interpretation. Its bound requires pure smooth feedback and an additional reference-comparison condition that has not been established for our frozen-weight local updates. Moreover, only the regret contribution necessarily vanishes; the comparison-error and reference-risk terms require additional control. Transferring this result to hard EO/EOD would also require controlling the hard–smooth discrepancy. We have explicitly distinguished this conditional result from guarantees for the implemented algorithm in the revision.

To examine the practical effect of multiple local updates, we additionally evaluated 1, 5, 10, and 20 local epochs on Adult, Dutch, and Credit Card, using three seeds per setting and 100 communication rounds. The results are as follows. These observations characterize the sensitivity to local training.

Table: EOD gaps under different local epochs.

Dataset1 epoch5 epochs10 epochs20 epochs
Adult0.070 (0.040)0.088 (0.067)0.095 (0.064)0.106 (0.037)
Credit Card0.067 (0.038)0.042 (0.023)0.038 (0.009)0.033 (0.029)
Dutch0.070 (0.027)0.059 (0.006)0.061 (0.010)0.061 (0.008)
AuthorsOpenReview ↗

[W2, Q3, Q4] Experiments on different classifiers, image datasets and partial training participation

Thank you for highlighting this limitation. We have conducted additional experiments addressing different classifiers, image data, and partial training participation. We consider five classifiers for examination, including MLP, Soft Decision Tree (SDT), SVM, GLM and CNN. An image dataset called CelebAMask-HQ is also introduced. The experiment results under EOD target are as follows.

Table: Test accuracy for different classifiers under the EOD target.

Bold marks the best unrounded mean within each dataset–classifier pair; -- denotes an unevaluated combination.

DatasetMethodMLPSDTLinear SVMGLMCNN
AdultFedAvg0.788 (0.032)0.803 (0.034)0.786 (0.061)0.804 (0.035)--
AdultFedFB0.793 (0.028)0.806 (0.031)0.799 (0.037)0.806 (0.031)--
AdultGloReFair0.803 (0.018)0.815 (0.010)0.774 (0.028)0.820 (0.010)--
Credit CardFedAvg0.796 (0.019)0.792 (0.010)0.780 (0.001)0.792 (0.010)--
Credit CardFedFB0.795 (0.018)0.789 (0.008)0.779 (0.000)0.792 (0.009)--
Credit CardGloReFair0.786 (0.007)0.784 (0.001)0.779 (0.000)0.788 (0.005)--
DutchFedAvg0.799 (0.024)0.806 (0.024)0.788 (0.028)0.806 (0.024)--
DutchFedFB0.658 (0.037)0.738 (0.050)0.738 (0.124)0.730 (0.062)--
DutchGloReFair0.627 (0.128)0.827 (0.002)0.690 (0.189)0.824 (0.003)--
CelebAMask-HQFedAvg--------0.889 (0.012)
CelebAMask-HQFedFB--------0.833 (0.101)
CelebAMask-HQGloReFair--------0.888 (0.011)

Table: Test EOD gap for different classifiers under the EOD target.

Bold marks the best unrounded mean within each dataset–classifier pair; -- denotes an unevaluated combination.

DatasetMethodMLPSDTLinear SVMGLMCNN
AdultFedAvg0.365 (0.298)0.357 (0.284)0.432 (0.395)0.351 (0.286)--
AdultFedFB0.325 (0.271)0.290 (0.188)0.312 (0.250)0.289 (0.169)--
AdultGloReFair0.236 (0.051)0.097 (0.085)0.145 (0.125)0.092 (0.060)--
Credit CardFedAvg0.183 (0.278)0.145 (0.146)0.010 (0.013)0.144 (0.153)--
Credit CardFedFB0.218 (0.270)0.115 (0.109)0.003 (0.003)0.134 (0.121)--
Credit CardGloReFair0.037 (0.040)0.027 (0.013)0.000 (0.000)0.031 (0.014)--
DutchFedAvg0.478 (0.207)0.477 (0.104)0.400 (0.337)0.476 (0.107)--
DutchFedFB0.328 (0.218)0.026 (0.016)0.104 (0.020)0.044 (0.047)--
DutchGloReFair0.027 (0.032)0.063 (0.009)0.063 (0.054)0.061 (0.006)--
CelebAMask-HQFedAvg--------0.112 (0.059)
CelebAMask-HQFedFB--------0.026 (0.019)
CelebAMask-HQGloReFair--------0.072 (0.020)

Regarding partial training participation, global group fairness in our setting is defined over the pooled population of all clients. Accordingly, GloReFair requires complete fairness feedback in each round. Directly pooling incomplete feedback computes group-conditional rates over the reporting clients’ observations, changing the population represented by the empirical fairness measurement. Our protocol therefore requires every client to evaluate the same aggregated model and return its group–label statistics, including clients omitted from local training. To demonstrate the performance of GloReFair under partial training participation, we evaluate all three tabular datasets with 10 clients, uniformly sampling 50%–100% of clients for training each round.

Table: Test accuracy for different fractions of clients missing from local training.

DatasetMethod0.00.10.20.30.40.5
AdultFedAvg0.800 (0.036)0.803 (0.036)0.796 (0.030)0.808 (0.017)0.764 (0.047)0.805 (0.030)
AdultFedFB0.804 (0.034)0.805 (0.033)0.798 (0.031)0.811 (0.020)0.779 (0.033)0.809 (0.030)
AdultGloReFair0.820 (0.008)0.815 (0.004)0.802 (0.009)0.804 (0.031)0.751 (0.099)0.813 (0.015)
Credit CardFedAvg0.786 (0.009)0.786 (0.010)0.786 (0.008)0.788 (0.014)0.784 (0.008)0.785 (0.010)
Credit CardFedFB0.786 (0.009)0.785 (0.009)0.786 (0.007)0.787 (0.012)0.785 (0.009)0.785 (0.009)
Credit CardGloReFair0.786 (0.005)0.788 (0.006)0.787 (0.006)0.789 (0.017)0.785 (0.009)0.785 (0.009)
DutchFedAvg0.807 (0.044)0.805 (0.043)0.811 (0.038)0.801 (0.027)0.804 (0.019)0.769 (0.022)
DutchFedFB0.812 (0.019)0.791 (0.013)0.798 (0.039)0.756 (0.107)0.734 (0.092)0.738 (0.049)
DutchGloReFair0.825 (0.002)0.820 (0.006)0.812 (0.019)0.786 (0.046)0.805 (0.032)0.767 (0.045)

Table: Test EOD gap for different fractions of clients missing from local training.

DatasetMethod0.00.10.20.30.40.5
AdultFedAvg0.467 (0.315)0.435 (0.291)0.439 (0.298)0.402 (0.255)0.355 (0.399)0.516 (0.264)
AdultFedFB0.370 (0.238)0.345 (0.222)0.364 (0.242)0.277 (0.176)0.281 (0.360)0.458 (0.206)
AdultGloReFair0.079 (0.060)0.075 (0.053)0.103 (0.080)0.094 (0.011)0.186 (0.129)0.181 (0.137)
Credit CardFedAvg0.118 (0.173)0.101 (0.156)0.087 (0.113)0.102 (0.169)0.059 (0.099)0.121 (0.207)
Credit CardFedFB0.093 (0.134)0.081 (0.116)0.077 (0.088)0.083 (0.124)0.067 (0.106)0.107 (0.180)
Credit CardGloReFair0.029 (0.005)0.031 (0.008)0.026 (0.007)0.021 (0.022)0.038 (0.050)0.065 (0.102)
DutchFedAvg0.383 (0.257)0.382 (0.252)0.404 (0.261)0.395 (0.239)0.369 (0.289)0.421 (0.246)
DutchFedFB0.059 (0.026)0.042 (0.022)0.036 (0.018)0.052 (0.016)0.030 (0.022)0.024 (0.013)
DutchGloReFair0.055 (0.012)0.078 (0.030)0.051 (0.029)0.093 (0.054)0.063 (0.022)0.068 (0.063)
Reviewer_ZQ26OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes fairness-aware reweighting method for federated learning setting, called GloReFair\texttt{GloReFair}, for global equal-opportunity and equalized-odds fairness under heterogeneous clients. The authors coined and used a server-side scalar, i.e., channel prices, that is used to turn collected fairness residuals into closed-form local sample weights. Several favorable properties, e.g., invariance, one-step centralized equivalence, and controller properties, and evaluate the method on three tabular benchmarks.

Critical strengths

The authors cleanly separate global fairness feedback from client-side optimization. Moreover, the closed-form weighting rule is favorable for practical FL setting and straightforward to implement. The empirical evaluation is extensive across three heterogeneity types, five severity levels, eight baselines with ablations.

Critical weaknesses
  • The main concern of the novelty is that the current manuscript is perceived as changing FedFB\texttt{FedFB}'s channel-level reweighting into sample-level reweighting.
    • In some cases, GloReFair showed comparative or even a bit lower performances relative to that of FedFB. I think substantial justification is needed that should favor sample-level reweighting rather than channel-level.
  • The proposed method relies on assumptions that may not hold in real-world FL systems.
    • It assumes full participation and exact pooled group-label statistics in every round.
    • We need analyses on how missing, delayed, noisy, or privately perturbed statistics affect the controller.
  • There are gaps in theoretical guarantee and the implemented algorithm.
    • Its centralized equivalence covers one synchronized full-batch step, but the actual method performs multiple local SGD steps.
  • The empirical design could have been extended to standard FL settings
    • While I fully appreciate extensive number of experiments, the overall design still uses logistic regression and tabular datasets.
    • Can it be done using practical and high-dimensional benchmarks tailored for FL, e.g., FLamby or LEAF?
  • The group-label statistics impose privacy concerns that should have been analyzed, although authors already accepted as one of their limitations.
  • The authors disclosed one of their use of AIs as "vibe coding", which casts doubts on the reproducibility and the integrity of the empirical results. How can it be confirmed and double-checked if it is vibe-coded?
Questions for authors
  • Please consider adding in-depth comparison (table or paragraph or both) between the GloReFair and FedFB (+ optionally Function Tracking paper), to clearly distinct the novelty and the motivation (i.e., when and why the proposed method should be preferred over existing methods) of the proposed method in federated settings?
  • How does partial participation affect pooled residuals, controller stability and resulting fairness?
  • Will there be any trade-offs or missing benefits when extending the analysis into local SGD setting (i.e., number of local steps > 1)?
11 comments
AuthorsOpenReview ↗

[W1, Q1] Motivation for sample-level reweighting.

Thank you for pointing out the motivation for sample-level reweighting and its relationship to FedFB. We agree that FedFB remains competitive in several settings, and that the justification should explain when the additional weighting flexibility is valuable.

FedFB applies a shared reweighting coefficient within each group–label channel. GloReFair combines a shared channel price with each sample’s current loss. The channel price is updated using globally pooled EO/EOD residuals, while the sample loss is evaluated locally. This gives the method an additional way to adjust relative training contributions within a channel. Samples sharing the same sensitive group and label can have different losses, including across heterogeneous clients. Loss-dependent weights allow their training emphasis to reflect this variation while retaining a common global fairness signal.

Our contribution concerns this coordination between global fairness feedback and local sample allocation. As discussed in Section 2, the threshold-based weighting form builds on Chai and Wang (2022). GloReFair connects this form to shared group–label prices driven by pooled EO/EOD residuals and studies the resulting mechanism under heterogeneous federated partitions.

The existing results provide evidence for the practical value of the complete method. Supplement H reports the comparative results under the strongest joint heterogeneity setting. On Adult, GloReFair achieves both higher mean accuracy and a smaller mean EOD gap. On CreditCard, the EOD gap decreases from 0.134 to 0.039 with an accuracy difference of 0.003. On Dutch, FedFB achieves the smaller EOD gap, while GloReFair achieves an accuracy improvement of 0.091. These comparisons motivate considering GloReFair when heterogeneous client data make maintaining predictive utility alongside global group fairness particularly important. We have clarified these mechanism differences and the conditions motivating their use in the revision.

AuthorsOpenReview ↗

[W2, Q2] Partial participation of clients

Thank you for your suggestions. We have completed additional experiments on Adult, Credit Card, and Dutch with 10 clients under joint Dirichlet heterogeneity (H = 0.2). In each round, we uniformly sample 10, 9, 8, 7, 6, or 5 clients without replacement for local training, corresponding to training-client absence rates of 0%–50%. Each setting uses three seeds, 100 communication rounds, 10 local epochs, and a batch size of 128.

The results are reported below. GloReFair achieves lower mean EOD gaps than FedAvg across all six participation levels on all three datasets, and lower mean EOD gaps than FedFB on Adult and Credit Card. These results indicate that GloReFair retains a global fairness advantage under partial training participation. We have added these experimental results to the revised appendix.

Table. Test accuracy for different fractions of clients missing from local training.

DatasetMethod0.00.10.20.30.40.5
AdultFedAvg0.800 (0.036)0.803 (0.036)0.796 (0.030)0.808 (0.017)0.764 (0.047)0.805 (0.030)
AdultFedFB0.804 (0.034)0.805 (0.033)0.798 (0.031)0.811 (0.020)0.779 (0.033)0.809 (0.030)
AdultGloReFair0.820 (0.008)0.815 (0.004)0.802 (0.009)0.804 (0.031)0.751 (0.099)0.813 (0.015)
Credit CardFedAvg0.786 (0.009)0.786 (0.010)0.786 (0.008)0.788 (0.014)0.784 (0.008)0.785 (0.010)
Credit CardFedFB0.786 (0.009)0.785 (0.009)0.786 (0.007)0.787 (0.012)0.785 (0.009)0.785 (0.009)
Credit CardGloReFair0.786 (0.005)0.788 (0.006)0.787 (0.006)0.789 (0.017)0.785 (0.009)0.785 (0.009)
DutchFedAvg0.807 (0.044)0.805 (0.043)0.811 (0.038)0.801 (0.027)0.804 (0.019)0.769 (0.022)
DutchFedFB0.812 (0.019)0.791 (0.013)0.798 (0.039)0.756 (0.107)0.734 (0.092)0.738 (0.049)
DutchGloReFair0.825 (0.002)0.820 (0.006)0.812 (0.019)0.786 (0.046)0.805 (0.032)0.767 (0.045)

Table. Test EOD gap for different fractions of clients missing from local training.

DatasetMethod0.00.10.20.30.40.5
AdultFedAvg0.467 (0.315)0.435 (0.291)0.439 (0.298)0.402 (0.255)0.355 (0.399)0.516 (0.264)
AdultFedFB0.370 (0.238)0.345 (0.222)0.364 (0.242)0.277 (0.176)0.281 (0.360)0.458 (0.206)
AdultGloReFair0.079 (0.060)0.075 (0.053)0.103 (0.080)0.094 (0.011)0.186 (0.129)0.181 (0.137)
Credit CardFedAvg0.118 (0.173)0.101 (0.156)0.087 (0.113)0.102 (0.169)0.059 (0.099)0.121 (0.207)
Credit CardFedFB0.093 (0.134)0.081 (0.116)0.077 (0.088)0.083 (0.124)0.067 (0.106)0.107 (0.180)
Credit CardGloReFair0.029 (0.005)0.031 (0.008)0.026 (0.007)0.021 (0.022)0.038 (0.050)0.065 (0.102)
DutchFedAvg0.383 (0.257)0.382 (0.252)0.404 (0.261)0.395 (0.239)0.369 (0.289)0.421 (0.246)
DutchFedFB0.059 (0.026)0.042 (0.022)0.036 (0.018)0.052 (0.016)0.030 (0.022)0.024 (0.013)
DutchGloReFair0.055 (0.012)0.078 (0.030)0.051 (0.029)0.093 (0.054)0.063 (0.022)0.068 (0.063)
AuthorsOpenReview ↗

[W3, Q3] Extending analysis into local SGD setting

We agree that the centralized equivalence applies only to one synchronized full-batch gradient step. For multiple local updates, Supplement C bounds the frozen-weight objective mismatch in terms of local model drift, providing a partial characterization of the effect of local training.

We further evaluated 1, 5, 10, and 20 local epochs across three datasets, using three seeds and 100 communication rounds. The final EOD gaps are reported below. Additionally, we visualize the training trajectories for different numbers of local epochs in Appendix I of our revised manuscript.

Table: EOD gaps under different numbers of local epochs.

Dataset1 epoch5 epochs10 epochs20 epochs
Adult0.070 (0.040)0.088 (0.067)0.095 (0.064)0.106 (0.037)
Credit Card0.067 (0.038)0.042 (0.023)0.038 (0.009)0.033 (0.029)
Dutch0.070 (0.027)0.059 (0.006)0.061 (0.010)0.061 (0.008)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.