acceptodds
Withdrawn from ICLR 2027

EquiConfig: Efficient Mitigation of Disparate Impact in Differentially Private Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Differentially private stochastic gradient descent (DP-SGD) enables neural network training with record-level privacy by clipping per-example gradients and adding calibrated noise to their aggregate. These operations can disproportionately weaken underrepresented groups' contributions, reducing their accuracy. Mitigation strategies rely on adaptive clipping and introduce extra parameters, increasing the computational and privacy costs of calibration. We propose EquiConfig, which screens DP-SGD's existing batch-size, clipping, and noise settings offline to construct a targeted shortlist for private calibration. Privacy accounting pairs each batch size with the noise required at the target privacy budget, forming a privacy ridge. Along each ridge, larger batches can strengthen group contributions relative to noise but leave fewer optimization steps under fixed privacy and epoch budgets. To assess this trade-off, EquiConfig defines a cumulative group-information score that combines per-update group signal relative to noise with expected group participation over training. Privacy-frontier elasticity analysis identifies batch increases that provably improve this score for every target group, while minimum total-update and group-activity requirements exclude configurations with too few training opportunities. We apply the screen to candidate grids spanning Gaussian, Laplace, and mixture-Laplace noise and multiple clipping bounds, yielding alternative configurations without adaptive training controls or gradient checkpoints. We compare against state-of-the-art mitigation methods both without private calibration and with DP-HyPO accounting for selection costs. Across vision, tabular, and language tasks, the results demonstrate substantial improvements in accuracy–disparity trade-offs at matched privacy targets, with fewer candidate evaluations under private calibration.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Disparate Impact, DPSGD, Differential Private Learning, Disparity Mitigation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Vt8WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the problem of reducing group-level accuracy disparities in differentially private learning without introducing additional adaptive clipping mechanisms or privacy-consuming calibration. Its main idea is the combination of an offline configuration-screening method with a group-information score.

Critical strengths

S1. The paper considers important DP-SGD parameters jointly and in particular recognizes that batch size, clipping, noise, and learning rate jointly influence training dynamics and privacy.

S2. The proposed cumulative group-information score, together with the PAN test, provides a solid and systematic way to compare configurations under fixed privacy and epoch budgets.

Critical weaknesses

W1. My main concern is about methodological novelty. The proposed EquiConfig primarily screens existing DP-SGD configurations using an analytically constructed score before applying a conventional privacy-accounted validation procedure. I do not see this a fundamentally new framework. The authors should distinguish their contributions from existing privacy-aware hyperparameter optimization, analytical configuration selection, and fairness-aware DP-SGD tuning methods.

W2. The paper's motivation is that DP-SGD may disproportionately harm underrepresented groups, and EquiConfig seeks configurations that improve the accuracy–disparity trade-off. However, DP-SGD does not necessarily impose a disproportionately larger accuracy loss on minority groups. Whether it does so depends on the dataset, model architecture, group characteristics, privacy budget, and training hyperparameters. The accuracy drop of minority group due to DP-SGD (compared with SGD) could be similar or even less than majority group. The paper conducts evaluation over MNIST, CIFAR-10, SST-2, QNLI, CelebA and Adult, some of which were modified from their original versions. I wondered how we know for sure the chosen dataset, model, and the hyperparameters incurred the disparate impact.

Questions for authors

Please see W2.

Reviewer_eTK1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a way to screen hyperparameter configurations to be used for hyperparameter optimization. The method is able to discard suboptimal hyperparameter combinations, reducing the search space, and leads to substantive improvements in fairness and utility metrics. The method is based on the PAN criterion derived by the authors that gives a score of how favorable the combination of batch sizes and clipping bounds are in terms of utility and fairness.

Critical strengths
  1. The empirical results strongly demonstrate the effectiveness of the method, in terms of fairness, but also in terms of utility.
  2. Unlike many other DP fairness methods, this method does not introduce additional hyperparameters.
  3. The privacy-frontier elasticity analysis (PAN) criterion is novel to me.
Critical weaknesses
  1. The method uses group counts to calculate PAN. But the group counts are private information.
  2. Table 1 mentions "Bounded-Adaptive" method, but is is not included in the baselines.
  3. The experiments mostly use single privacy budget ε\varepsilon per dataset. Preferably, to demonstrate the effectiveness of the method, the paper should have a range of ε\varepsilon values for each dataset.
  4. Some hyperparameter choices seem dubious (small number of epochs, not using very large batches, etc). More details in the questions section.
  5. There are no confidence intervals for the results.
Questions for authors
  1. In Figure 4 it seems that the overall accuracy is better than when training from scratch for \varpepsilon=1\varpepsilon=1. Why this might be?
  2. Figure 4 is an image. Can you provide a LaTex Table version of it?
  3. I am very interested how much of the improvements are due to the noise mechanism. Could you do an ablation that only include a single noise mechanism at time to compare the results? If that is not possible, an ablation using only the Gaussian mechanism would be appreciated.
  4. The number of epochs seems really low. Have the models converged?
  5. The common hyperparameters (Table 9) do include only very small learning rates. Have you checked that selected parameters are not at the edge of your grid?
  6. The batch sizes also do not seem to include very large batch sizes, but earlier you state that "larger batches may improve group signal". If the issue is that the model has no time to converge due to the low numbers of steps, could you arrange an experiment where you also increase the number of epochs along with batch sizes?
  7. The different image datasets seem to use different models. How did you justify these selections? The same models for multiple datasets would be more systematic and compelling.
  8. (minor) Some of the paragraphs are very long and hard to read, e.g. The "Exhaustive Search" paragraph of Section 4.2. This makes them hard to read and understand. Writing could be improved.
Reviewer_z8ykOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EquiConfig is a screening step that runs before private hyperparameter search for DP-SGD. The goal is to improve accuracy/disparity trade-offs for minority groups by choosing batch size, clipping and noise family (Gaussian, Lap2, PLRV-O) well, instead of adding adaptive-clipping machinery as DPSGD-F or DPSGD-G-A do. For each noise family and batch size the accountant is inverted to get the noise scale at the target budget. A "group information" score (Eq. 4) multiplies the expected number of updates in which a group appears by a squared signal-to-noise proxy. The PAN test (Theorem 1) keeps adjacent batch sizes where that score goes up for every target group, floors on total and group-active update counts remove infeasible points, and within-branch Pareto pruning produces the shortlist. DP-HyPO then searches the shortlist with private validation. Experiments: imbalanced MNIST, CIFAR-10 (TinyViT), SST-2 and QNLI (BERT, RoBERTa prompt tuning), Adult and CelebA, against DP-SGD, DPSGD-F and DPSGD-G-A, both with exhaustive search and with private calibration. The appendices derive RDP bounds for Lap2 and a coordinatewise PLRV-O, asymptotics of the score, and DP-HyPO horizon results.

Critical strengths

EquiConfig is a screening step that runs before private hyperparameter search for DP-SGD. The goal is to improve accuracy/disparity trade-offs for minority groups by choosing batch size BB, clipping bound CC and noise family (Gaussian, Lap2, PLRV-O) well, instead of adding adaptive-clipping machinery as DPSGD-F or DPSGD-G-A do. For each noise family jj and batch size the accountant is inverted to get the noise scale sj,⋆(B)s_{j,\star}(B) at the target (εtrain,δtrain)(\varepsilon_{\mathrm{train}}, \delta_{\mathrm{train}}). A "group information" score Eq. 4) multiplies the expected number of updates in which group ν\nu appears by a squared signal-to-noise proxy. The PAN test (Theorem 1) keeps adjacent batch sizes where that score goes up for every target group, floors T(B)≥Tmin⁡T(B) \ge T_{\min} and Aν(B)≥Aν,min⁡A_\nu(B) \ge A_{\nu,\min} remove points with too few updates, and within-branch Pareto pruning produces the shortlist. DP-HyPO then searches the shortlist with private validation. Experiments: imbalanced MNIST, CIFAR-10 (TinyViT), SST-2 and QNLI (BERT, RoBERTa prompt tuning), Adult and CelebA, against DP-SGD, DPSGD-F and DPSGD-G-A, both with exhaustive search and with private calibration. The appendices derive RDP bounds for Lap2 and a coordinatewise PLRV-O, asymptotics of the score, and DP-HyPO horizon results.

Critical weaknesses
  1. The headline results do not measure the thing the paper proposes. Table 12 shows that under exhaustive search every ablation (no PAN, no floors, no Pareto) lands on exactly the same accuracy and disparity. So the screen never changes the optimum, it only shrinks the candidate list. But Tables 2, 3, 4 and Fig. 4 present the exhaustive-search column as EquiConfig beating DP-SGD. The DP-SGD baseline is searched over Gaussian noise only (Table 9) while EquiConfig's grid adds PLRV-O noise with its own clipping values, so the exhaustive gains must come from the larger search space, most likely from the non-Gaussian noise, which is prior work (Yang et al. 2025). The paper never says which configuration won. The one piece of evidence for the screen itself is the private-calibration ablation in Table 12, on one dataset at one budget. The baselines also look mis-tuned: DPSGD-F and DPSGD-G-A are swept over C∈{1,10,20}C \in \{1, 10, 20\} with two values per method-specific parameter; DPSGD-G-A ends up with worse disparity than plain DP-SGD in most settings (38% on MNIST, 62% on QNLI), the opposite of what its own paper reports, and the SST-2 baselines in Fig. 4 collapse to 3% and 0% minority accuracy. The authors acknowledge the discrepancy but do not resolve it. Bounded adaptive clipping (Zhao et al. 2026) appears as a competitor in Table 1 and is then never run.

  2. I cannot check the private-calibration claims, and there is no variance anywhere. Appendix F.1 does not say how validation accuracy and disparity are privatized (mechanism, sensitivity, noise, per-query ε\varepsilon) or how ρω(α)\rho_\omega(\alpha) is built, so εhypo∈{6.6,13.9,11.4,1.4}\varepsilon_{\mathrm{hypo}} \in \{6.6, 13.9, 11.4, 1.4\} cannot be verified. Eq. 9 says select by minimum disparity under an accuracy gate U(fω)≥Umin⁡U(f_\omega) \ge U_{\min}; Appendix F.3.2 says the score is U=0.5 Acc+0.5 (1−Disp)U = 0.5\,\mathrm{Acc} + 0.5\,(1 - \mathrm{Disp}). No table reports seeds or intervals even though DP-HyPO is a randomized search, and the numbers move a lot: EquiConfig's MNIST private-calibration result is 93.56 / 28.81 in Table 2 (εhypo=6.6\varepsilon_{\mathrm{hypo}} = 6.6) and 97.3 / 10.9 in Table 12 (εhypo=7.5\varepsilon_{\mathrm{hypo}} = 7.5), with the same εtrain=3\varepsilon_{\mathrm{train}} = 3. There are plain inconsistencies as well: exhaustive accuracy 97.85 in Table 2 vs 97.45 in Table 12; Section 4.2 quotes DP-SGD at 91% / 47% where Table 2 says 92.43 / 39.96, and compares EquiConfig's CIFAR private-calibration numbers against DP-SGD's exhaustive ones; QNLI εtrain\varepsilon_{\mathrm{train}} is 12 in Table 8 and 20 in Table 4.

Questions for authors
  1. Which configuration (noise family, BB, CC, η\eta) was selected in each reported result? And can you add DP-SGD searched exhaustively over the same grid, PLRV-O included? Without that I cannot tell what the screen contributes. Please spell out the private evaluator: how accuracy and disparity are privatized, the per-query cost, how ρω(α)\rho_\omega(\alpha) is formed, and where metadata and reference costs go. Which objective was actually used, Eq. 9 or the F.3.2 score?

  2. How many runs are behind Tables 2-4, 10-12 and Figs. 4-6? Mean and std over seeds, including the DP-HyPO randomness, are needed. Why do Table 2 and Table 12 disagree?

  3. Are the group counts nνn_\nu treated as public? If not, what did privatizing them cost and where is it charged? Did you reproduce the original DPSGD-F / DPSGD-G-A numbers under their own protocol before putting them on the wider grid? Why is bounded adaptive clipping missing?

  4. Was Lap2 actually in the executed grids? Table 9 only lists Gaussian and PLRV-O for the vision tasks. And what are the exact Rehn et al. pre-filter rules from Appendix F.3.1?

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

Related papers

Discussion (0)

Sign in to comment.