acceptodds
Under review as a conference paper at ICLR 2027

Protecting the Group, Not Just Its Members: Feasible Differential Privacy for Groups

Abstract

Large amounts of data are required to infer robust statistics or train large machine-learned models. Not only a public release of such data can reveal sensitive information of the individuals, but such information can also be inferred from a statistical estimator or learned predictor. Differential privacy (DP) provides strong guarantees for protecting individual records; however, these guarantees do not prevent sensitive properties of groups from being inferred. In many applications including healthcare, protecting the characteristics of a subpopulation within an institution can be as important as protecting individual records, for example disease prevalence within a demographic group at a hospital. We introduce *Differentially Private Attributes of Groups (DPAG)*, a framework that treats such groups as the atomic units of privacy. Rather than obtaining group protection by scaling an individual-level privacy guarantee with group size, DPAG directly bounds and perturbs each protected group's contribution. We develop mechanisms for centralized learning and extend the framework to cross-silo federated learning that is regularly used in healthcare scenarios to transfer institutional knowledge for machine learning without sharing the underlying data. We evaluate DPAG on centralized and federated clinical data, comparing direct group-level protection with standard DP mechanisms and conventional group-privacy approaches. Our results show that **DPAG can make group-level protection feasible without a privacy cost that grows with the number of individuals in the group, while retaining useful model utility**.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
differential privacy, federated learning, privacy

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_XARiOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The author(s) propose differentially private attributes of groups (DPAG) as a framework to ensure group-level privacy that directly bounds the contribution of each client-group pair towards training of machine learning models. DPAG avoids the disadvantage of extending individual-level DP to group privacy, where the privacy parameter ϵ\epsilon scales with the number of individuals in the group. It is primarily intended to improve the group-level privacy–utility trade-off in cross-silo federated learning, but can also be extended to centralized learning settings. It achieves this by regarding the client–group pair as the atomic unit of privacy, bounding its contribution, and calibrating noise accordingly. Tables 1 and 2 demonstrate utility benefits on the evaluated tasks: DPAG avoids the ineffective group-privacy conversion (as in the case of standard individual-level DP) in the centralized setting and achieves substantially higher utility than client-level DP at matching (cell-level) privacy budgets in the FL setting.

Critical strengths
  • The theory part of the paper is relatively straightforward. The author(s) primarily change the atomic unit of privacy from an individual to a cell, where a cell represents a client–group pair protected under DP, and subsequently develop contribution-bounding and noise-calibration mechanisms for this privacy unit.
  • In the federated setting, DPAG avoids the group-size penalty of lifting standard individual-level DP to group privacy (the privacy budget ϵ\epsilon scales by the number of individuals in the group). For a fixed clipping bound, DPAG reduces sensitivity by a factor of ∣G∣|G| compared with client-level DP, where ∣G∣|G| is the number of predefined groups.
  • The cell-based privacy audit using a group-level analogue of membership inference (Appendix A.6) strengthens the paper's position. DPAG does preserve utility at the risk of higher attack AUC but client-level DP can severely impact the utility to preserve privacy. This is a candid analysis undertaken by the author(s) rather than trying to embellish the paper with selective (positive) results in favor of DPAG.
Critical weaknesses
  • In the FL setting, the paper assumes roughly equal group sizes/ distribution among clients. It does not systematically investigate how group-size imbalance across clients affects utility, particularly for small groups. The authors acknowledge the related challenge of learning from smaller group-specific datasets in lines 472–474 (Section 6; Limitations).
Questions for authors

Suggestions

  • Neither the abstract nor the contributions mention any results which quantify the advantage of using DPAG over the direct cell-level standard DP or group privacy algorithms. This is needed to put into perspective the gain one should expect if they switch to DPAG as opposed to the other alternatives.

  • With the contributions listed in Page 2, also mention the sections in the paper corresponding to them. This enhances readability of the paper.

  • The wrapfigure for Figure 1 needs to be corrected. Blank space below it is leaking into Page 6.

5 comments
AuthorsOpenReview ↗

Direct Cell-Level Comparison and Imbalanced-Group Analysis Strengthen the Empirical Evaluation

Thank you for the positive assessment and for highlighting the practical relevance of the privacy objective, the 1/∣G∣1/|G| sensitivity reduction in the federated mechanism, and the complete-cell replacement audit. We agree that the two points you raise can make the empirical contribution substantially clearer, we address them below.

Group-size imbalance - To clarify, the real-world imaging experiment in fact already operates in a strongly imbalanced regime. The ten group slots per hospital are fixed and public, but only 91 of the 100 possible hospital-group cells are non-empty. Among non-empty cells, the median size is 3 patients and the maximum is 33, with several hospitals missing particular groups entirely. The formal sensitivity result is independent of these counts because every public slot receives the same fixed weight 1/∣G∣1/|G|. We agree with your broader point that very small groups can still reduce utility because their cell-specific trajectories must be estimated from little data. We will make this distinction clearer in the discussion.

Direct cell-level DP comparison - We also followed your suggestion to quantify the benefit relative to a standard mechanism operating at exactly the same privacy granularity. We implemented DP-FedAvg in which every hospital-group cell is treated as an independent virtual user, while matching DPAG's protected cells, contribution cap, clipping norm, fixed aggregation weights, privacy budgets, model, and optimization.

At ϵ=8,16,32\epsilon=8,16,32, this direct cell-level baseline obtains Dice

0.186\pm0.091,\quad 0.230\pm0.185,$$ whereas DPAG obtains $$0.631\pm0.030,\quad 0.692\pm0.043,\quad 0.730\pm0.030.$$ This comparison shows the benefit of the federated construction on top of changing the privacy unit. If the $HG$ cells are privatized independently before server aggregation, the aggregate Gaussian noise scale is $2Sz/\sqrt{HG}$. DPAG instead first bounds and aggregates the $G$ cell contributions within each institution and privatizes one institutional message with cell-replacement sensitivity $2S/G$, yielding aggregate noise $2Sz/$G\sqrt H\$\$. Hence DPAG reduces the aggregate noise by a factor $\sqrt G$ with $G=10$ in our experiments. With the existing client-update comparison, this shows that protecting the whole institution is unnecessarily costly when the sensitive object is one hospital-group cell. Privatizing every cell separately also loses utility because it fails to exploit the nested cross-silo structure. DPAG, on the other hand, retains substantially higher utility while preserving the same formal cell-level adjacency. We will add this quantitative comparison to the abstract and contribution discussion. We will also add section references to the contribution bullets and correct the Figure 1 wrapping issue as suggested. We appreciate that the review identified the cell-replacement audit as a strength. We will retain its current cautious interpretation in that it provides empirical evidence complementary to the formal guarantee, while the formal privacy claim continues to come from the DP mechanism itself.

Dear Reviewer XARi,

Thank you again for the positive assessment and constructive suggestions. We have now added the direct cell-level DP-FedAvg comparison, clarified the existing group-size imbalance, and incorporated the presentation changes you suggested.

The new matched-granularity experiment further strengthens the empirical case for DPAG. Please let us know if any questions remain.

AuthorsOpenReview ↗

Discussion period ends today

Dear Reviewer XARi,

As the author discussion period ends today, we just wanted to briefly follow up and thank you again for your positive assessment and constructive suggestions.

We have incorporated the requested matched cell-level comparison and presentation clarifications. Please let us know if any questions remain before the discussion closes.

Reviewer_4j8aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Differentially Private Attributes of Groups (DPAG), a framework for protecting predefined groups, such as hospital–demographic cells, rather than individual records. The proposed mechanisms clip group-level contributions and add Gaussian noise calibrated to cell-level sensitivity. They apply these mechanisms to centralized learning and cross-silo federated learning, with experiments on clinical prediction and medical image segmentation tasks.

Critical strengths

The paper considers a practically relevant privacy objective: protecting the information associated with demographic subgroups within institutions. The authors provide experimentss on real clinical datasets, demonstrating that directly protecting these cells can retain more utility than the chosen record-level and client-level DP baselines.

Critical weaknesses

1. Limited novelty relative to existing user-level DP methods.

The main contribution appears to be a straightforward application of standard user-level DP, treating each predefined group or hospital–demographic cell as the privacy unit. Clipping contributions at the privacy-unit level and adding Gaussian noise calibrated to their sensitivity are well-established techniques. The claimed improvement over conventional group-privacy scaling follows directly from defining the entire cell as a single protected unit. It is unclear what fundamentally new privacy definition, algorithm, or theoretical insight is introduced here.

The related-work discussion also omits relevant literature on privacy-preserving cross-silo FL without a trusted server, including Lowy and Razaviyayn (ICLR 2023), Lowy et al. (AISTATS 2023), and Gao et al. (ICML 2024). These works develop the inter-silo record-level DP (ISRL-DP) framework and corresponding algorithms and privacy–utility analyses. Although ISRL-DP targets individual records within silos rather than entire cells, these papers are relevant to positioning the proposed approach within the existing cross-silo FL literature.

2. Inadequate baselines to establish the claimed advantages.

The main empirical comparisons are against record-level DP converted to group privacy and client-level DP protecting entire institutions. These are not the most natural comparisons for assessing the proposed method, since they either use a different privacy unit or protect a strictly larger unit than necessary.

A more informative baseline would be standard user-level DP-SGD or DP-FedAvg with each hospital–demographic cell treated directly as a privacy unit. It is unclear how such a baseline would differ from DPAG, and without this comparison, the reported utility gains do not convincingly establish an advantage over existing methods. Instead, they largely demonstrate the expected benefits of calibrating privacy to the intended protection granularity.

Questions for authors
  1. What is the fundamental methodological difference between DPAG and standard user-level DP when each predefined cell is treated as a single privacy unit? Is there a new privacy guarantee or algorithmic ingredient beyond this choice of granularity?

  2. How would the proposed mechanisms compare against standard user-level DP-SGD or DP-FedAvg configured with the same protected cells, aggregation weights, and privacy budgets? Would these methods yield essentially the same sensitivity bounds and algorithms?

3 comments
AuthorsOpenReview ↗

DPAG’s Nested Cross-Silo Construction Reduces Sensitivity Beyond Virtual-User DP-FedAvg

We thank the reviewer for the constructive feedback. First, we would like to clarify that the matched-cell comparison requested in the review considers a different problem than the one in the current manuscript. More specifically, standard virtual-user DP-FedAvg and DPAG do not yield the same mechanism or sensitivity once the cross-silo release structure is taken into account. We further clarify this below

Methodological difference - We agree that DPAG, like many privacy methods, builds on standard Gaussian and RDP machinery. What is new in the algorithm is the ordering of contribution bounding, aggregation, and privatization when multiple protected privacy units reside inside one silo. In particular, we first conduct a cell-specific training, only then do cell clipping, followed by within-silo aggregation. Finally, a privatised release from each silo is done.

Importantly, with this mechanism for GG protected cells, replacing one cell changes only one clipped term in the institutional average, yielding message sensitivity 2S/G2S/G. This is not obtained by ordinary pooled local training with user-level clipping, whose sensitivity is 2S2S, nor by treating the cells as independently released private users.

Same-cell DP-FedAvg baseline - Following the review, we implemented exactly this latter baseline. Each hospital-group cell is treated as a virtual DP-FedAvg user, while matching DPAG's protected cells, contribution cap, clipping norm, fixed aggregation weights, privacy budgets, model, and optimization. Each virtual-cell message is independently privatized before aggregation.

From this experiment, we get Dice scores at ϵ=8,16,32\epsilon=8,16,32 of 0.157±0.0850.157\pm0.085, 0.186±0.0910.186\pm0.091, and 0.230±0.1850.230\pm0.185, versus 0.631±0.0300.631\pm0.030, 0.692±0.0430.692\pm0.043, and 0.730±0.0300.730\pm0.030 for DPAG.

The difference can be seen already analytically. For HH institutions and GG protected cells per institution, independently privatizing HGHG virtual users gives aggregate noise

σvirtual=2SzHG.\sigma_{\mathrm{virtual}}=\frac{2Sz}{\sqrt{HG}}.

DPAG instead releases HH institutional messages, each having cell-replacement sensitivity 2S/G2S/G, giving

σDPAG=2SzGH.\sigma_{\mathrm{DPAG}}=\frac{2Sz}{G\sqrt H}.

Hence the matched virtual-user baseline incurs a factor G\sqrt G more aggregate noise. In our experiment we have G=10G=10, so the factor is 10≈3.16\sqrt{10}\approx3.16. The empirical results reflect this analytical difference.

We will revise the novelty statement accordingly. We do not claim a new DP accountant or noise distribution. The technical contribution is a cell-private federated mechanism for a nested privacy unit, including the cell-specific contribution bounding before within-silo aggregation and the resulting 2S/G2S/G sensitivity under arbitrary cell replacement. The new matched-cell experiment shows that this mechanism has a substantial utility consequence beyond simply comparing against a stronger client-level privacy unit.

Related work - Thank you for this suggestion. We will happily add Lowy and Razaviyayn (ICLR 2023), Lowy et al. (AISTATS 2023), and Gao et al. (ICML 2024). These works protect records within silos rather than arbitrary replacement of an entire hospital-group cell, but they are indeed relevant to positioning the trust model and cross-silo setting.

Dear Reviewer 4j8a,

We would like to briefly follow up on our rebuttal. Following your suggestion, we added a matched cell-level DP-FedAvg baseline in which every hospital--group cell is treated as a virtual user under the same privacy budgets, clipping, aggregation weights, model, and training protocol.

The comparison directly addresses your question of whether this yields essentially the same sensitivity and algorithm as DPAG: the release structures differ, and independently privatizing the cells incurs a factor ∣G∣\sqrt{|G|} more aggregate noise than DPAG's bounded within-silo aggregation. The new experiments show a substantial corresponding utility gap.

We also added the suggested ISRL-DP literature and clarified the methodological positioning.

We hope this addresses the main novelty and baseline concerns in your review, and we would appreciate your reconsideration of the evaluation. We would be happy to clarify any remaining questions.

AuthorsOpenReview ↗

Discussion period ends today

Dear Reviewer 4j8a,

As the author discussion period ends today, we wanted to follow up once more and ask whether you could please let us know if our rebuttal and the new matched cell-level baseline have addressed your concerns.

The new experiment directly tests the same privacy unit, aggregation weights, clipping, and privacy budgets that you requested, and we also clarified the methodological distinction from standard user-level DP and added the suggested ISRL-DP literature.

We would greatly appreciate a brief response before the discussion closes, particularly if any concerns remain that we can still address.

Reviewer_t6y6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Differentially Private Attributes of Groups (DPAG), which treats a predefined demographic or institution-group cell as the atomic privacy unit. DPAG clips and balances each cell's contribution before adding calibrated noise, avoiding the group-size dependence incurred when record-level DP is converted to group privacy. The framework is instantiated for centralized statistics, centralized learning, and cross-silo federated learning. Experiments on NHANES and UTSW-Glioma show that DPAG retains substantially more utility than record-level group-privacy conversion or client-update DP. The paper also reports a complete-cell replacement attack as an empirical privacy audit.

Critical strengths
  1. The paper defines a clear privacy objective between record-level and client-level privacy and provides straightforward sensitivity and composition analyses.

  2. The experiments include centralized and federated settings, replicated imaging results, structural ablations, and an empirical cell-replacement attack. The paper is also careful to distinguish attack-specific evidence from formal DP guarantees.

Critical weaknesses
  1. The technical novelty is limited. DPAG largely treats each cell as a user-level privacy unit and applies standard clipping and Gaussian noise. The distinction from user-level, element-level, and attribute privacy needs to be sharper.

  2. The main comparisons use different privacy units. Client-update DP protects a stronger unit with sensitivity 2S2S, while DPAG protects a cell with sensitivity 2S/∣G∣2S/|\mathcal G|. The centralized cell-equivalent baseline also fails to achieve the target budget. In addition, ϵ∈8,16,32\epsilon\in{8,16,32} with δ=0.01\delta=0.01 provides relatively weak protection for a healthcare application.

Questions for authors
  1. What is technically new beyond treating each hospital-group cell as a user and applying standard user-level clipping and noise?

  2. Can the authors provide a valid baseline evaluated under the same cell-level adjacency and target (ϵ,δ)(\epsilon,\delta)?

  3. How does DPAG perform under smaller values of ϵ\epsilon and δ\delta, such as ϵ≤4\epsilon\leq4 and δ≪0.01\delta\ll0.01?

  4. Can the authors separate the effect of privacy noise from group-balanced training using a non-private cell-balanced baseline in the centralized experiment?

3 comments
AuthorsOpenReview ↗

Matched Cell-Level Baselines and Ablations Clarify DPAG’s Privacy–Utility Trade-off

Thank you for the constructive feedback, we additionally ran the two suggested comparisons.

Technical novelty - DPAG uses standard Gaussian noise and RDP accounting. The new part is the federated mechanism used to protect a cell when several protected cells reside inside one communicating institution. DPAG trains and clips the cell-specific trajectories separately, aggregates them within the institution using fixed public weights, and only then privatizes the institutional message. Under arbitrary replacement of one hospital-group cell, only one bounded term changes, giving sensitivity 2S/∣G∣2S/|G|. In contrast, whole-client clipping gives weakerweaker sensitivity 2S2S, while separately releasing the cells as private users leads to a different mechanism and noise scale. We will clarify this distinction from user-level, element-level, and attribute privacy.

Matched cell-level baseline - We implemented the suggested DP-FedAvg baseline in which every hospital-group cell is treated as a virtual user. We match the cells, clipping norm, contribution cap, fixed weights, privacy budgets, model, and training protocol. Each virtual-cell message is independently privatized before server aggregation. At ϵ=8,16,32\epsilon=8,16,32, this baseline obtains Dice 0.157±0.0850.157\pm0.085, 0.186±0.0910.186\pm0.091, and 0.230±0.1850.230\pm0.185, compared with 0.631±0.0300.631\pm0.030, 0.692±0.0430.692\pm0.043, and 0.730±0.0300.730\pm0.030 for DPAG.

This difference is given by the sensitivity analysis. With HH institutions and GG cells per institution, the virtual-user construction has aggregate noise standard deviation 2Sz/HG2Sz/\sqrt{HG}. DPAG first performs bounded within-institution aggregation and therefore has 2Sz/2Sz/G\sqrt H$$. The virtual-user construction is thus noisier by a factor G\sqrt G, whichhere is G=10G=10.

Privacy budgets - We agree that the evaluated imaging budgets should not be interpreted as deployment recommendations for healthcare. Our replicated imaging experiments currently cover RDP ϵ∈8,16,32\epsilon\in{8,16,32} at δ=0.01\delta=0.01 and have not run the same full sweep at substantially smaller ϵ\epsilon and δ\delta. We will state this explicitly in the discussion with respect to down-stream scenarios. Note that this does not affect the mechanism or the comparisons above, but only limits conclusions in substantially tighter privacy regimes.

Centralized no-noise ablation - We also added the suggested cell-balanced, clipped, and non-private baseline. Across three seeds it obtains AUROC 0.828±0.0050.828\pm0.005, AUPRC 0.500±0.0180.500\pm0.018, and Brier 0.143±0.0080.143\pm0.008. The corresponding values are 0.917/0.606/0.0620.917/0.606/0.062 for ordinary non-private training and 0.787/0.422/0.1580.787/0.422/0.158 for private DPAG. Thus part of the utility loss comes from cell-balanced and clipped training itself, and privacy noise causes an additional loss. We will add this result for a more complete picture.

Dear Reviewer t6y6,

We would like to briefly follow up on our rebuttal. In particular, we ran both additional comparisons suggested in your review: the matched cell-level DP-FedAvg baseline and the centralized cell-balanced no-noise ablation.

The matched baseline confirms that DPAG's gain is not only due to comparing against a stronger client-level privacy unit, while the no-noise ablation separates the structural training cost from the additional privacy-noise cost.

We hope these new results and our clarification of the methodological contribution address your main concerns, and we would appreciate your reconsideration of the evaluation. We would be happy to clarify any remaining questions.

AuthorsOpenReview ↗

Discussion period ends today

Dear Reviewer t6y6,

As the author discussion period ends today, we wanted to follow up once more and ask whether you could please let us know if our rebuttal and the additional experiments have addressed your concerns.

In particular, we added both the matched cell-level DP-FedAvg baseline and the centralized no-noise ablation that you requested, and clarified the methodological novelty and remaining privacy-budget limitation.

We would greatly appreciate a brief response before the discussion closes, especially if any concerns remain that we can still clarify.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.