acceptodds
Under review as a conference paper at ICLR 2027

Privacy Considerations in Multi-Institutional Data Harmonization Through a Category-Theoretic Lens

Abstract

Institutions pool data to overcome local scarcity, and filtering-based matching harmonizes data pooling by admitting only samples that align with a reference site. We show it is unsafe for privacy, and that the two failures have a common cause. For an institution/site whose population not harmonized with respect to other institutions, the samples that align best with the reference institution/site are precisely the atypical ones in its own distribution, so the attempt to harmonize data forces each site to contribute an unrepresentative subset of itself. What leaks is then not a patient record but the distributional identity of the institution. We therefore propose pooling specific a – DP metric () for data pooling, a task for which privacy has not been studied and is fundamentally different from existing DP setups. Using category theory helps us demonstrate and understand the natural tension between data harmonization and , proving a structural incompatibility where improving data harmonization (utility) strictly degrades (privacy) for confounded institutions under data distribution shifts. To resolve this, we propose privacy-budgeted minimax matching (PBMM) that jointly optimizes for distributional alignment, sample efficiency, and differential privacy. Across synthetic, multi-site fMRI (homogeneous institutional distribution) and six-modality medical imaging benchmarks (heterogeneous institutional distribution), PBMM preserves the harmonization of filtering while cutting peak in-domain exposure by over and recovering samples that filtering discards.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
privacy, data pooling, data harmonization, data distribution shift, data scaling

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_fBR6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that filtering-based pooling selects atypical samples from non-anchor sites, leaking institutional distribution rather than individual records. It proposes a pooling-specific DP notion, DPpoolDP_{pool}, and PBMM. Experiments cover synthetic data, ABIDE, six medical imaging sets, and fMRI.

Critical strengths
  1. Identifies a real privacy failure at sample selection.
  2. PBMM is simple, model-agnostic, and tested across domains.
Critical weaknesses
  1. Non-standard privacy notion. No standard DP guarantee or finite-budget results.
  2. Category theory is decorative. “Minimax” is misused.
Questions for authors
  1. Is the sensitivity for μ(Xksel)\mu(X_k^{\mathrm{sel}}) or μ(P)\mu(\mathcal{P})? Why no nksel/Nn_k^{\mathrm{sel}}/N factor?
  2. What ϵkmax⁡\epsilon_k^{\max} is used in main experiments?
  3. What is the multi-class objective? Eq. 6 and Algorithm 1 differ.
Reviewer_HD7FOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies privacy risks in data pooling across institutions with different data distributions. It argues that matching-based methods may select samples that are unusual for their original institution. To address this, the authors introduce (DPpoolDP_{\mathrm{pool}}), a privacy measure based on the difference between the selected data mean and the institution's full data mean. They also propose Privacy-Budgeted Minimax Matching, which combines matching, sample selection, and privacy budget checks. The method is evaluated across several synthetic and real-world settings. The results show lower measured privacy costs while maintaining competitive performance in most cases, although some experiments show reduced accuracy.

Critical strengths
  • Important problem. The paper highlights a useful concern with matching-based data pooling. Selecting samples that align with a target institution may result in selecting unusual samples from other institutions. Studying this effect while considering data utility and privacy is valuable, especially for medical data sharing.

  • Broad experimental evaluation. The authors evaluate PBMM on both synthetic and real datasets, including different medical imaging tasks. The method achieves lower measured privacy costs while maintaining competitive performance in several settings. These results suggest that the proposed sample selection method may have practical value.

Critical weaknesses
  • The claimed privacy guarantee is not established. Definition 2 measures privacy cost using the distance between the selected subset mean and the institution's full data mean. This does not generally measure how much the global pooled mean changes when an institution replaces its selected subset with its full dataset. The measure also assigns zero cost when an institution shares its entire dataset. Appendix A.1 acknowledges that DPpoolDP_{\mathrm{pool}} is not a standard worst-case DP guarantee. However, Lemma 3 still claims a formal privacy guarantee for the pooled mean. The proof assumes a noisy mean release, while Algorithm 1 returns selected samples without adding noise. It is therefore unclear what privacy protection the actual method provides. This affects the paper's main contribution.

  • The algorithm and theoretical guarantees are not consistent. Algorithm 1 begins by adding the matched core without first verifying that it fits within the privacy budget. This is problematic because the paper indicates that matching can incur large privacy costs for shifted institutions. Also, Algorithm 1 evaluates privacy costs for candidate bands prior to Stage 3, but it does not recompute these costs after each band is incorporated, bands that individually satisfy the budget constraint may jointly violate it once combined. This conflicts with the cumulative budget enforcement described in Section 3.2.3 and undermines Lemma 3. Another issue is that Lemma 1 treats the matched-core mean as a surrogate for the full institution mean, even though the paper's central motivation is that the matched core may substantially differ from the institution mean. Overall, authors do not adequately establish when this proxy is reliable in the difficult cases being studied.

  • One formatting issue is that the authors appear to have overused LaTeX \vspace{} (or similar commands), which causes text to be unevenly spaced and nearly overlapping in places. In its current form, the paper is not ready for publication.

Questions for authors

1- What exact privacy guarantee does Algorithm 1 provide for its actual output? Can the authors explain how Lemma 3 applies when the algorithm returns selected samples rather than a noisy pooled mean? How does the sensitivity calculation account for changes in the global pooled mean when the number of selected samples changes?

2- How does PBMM handle an institution whose initial matched core already exceeds the allowed privacy budget? Why are the privacy costs not recomputed after each band is added in Stage 3? Can the authors provide a corrected algorithm and a proof that the budget is always satisfied?

3- Lemma 1 uses the matched-core mean as a proxy for the full institution mean. However, the paper argues that matching selects unusual samples from strongly shifted institutions. How reliable is this proxy in such cases? Can the authors provide experimental results showing the proxy error for the most strongly shifted institutions?

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

Filtering-based pooling keeps a sample when it lies closer to its anchor class centroid than to any other class. For a site whose mass sits far from the anchor, the points that pass are the tail of that site. The authors define a pooling score, DPpool, by putting the gap between the admitted mean and the site mean into the Gaussian-mechanism formula, and they propose privacy-budgeted minimax matching: take the matched core, reweight toward that core, and add inner radial bands before outer ones until a per-site cap is hit. Experiments use synthetic Gaussians, ABIDE, six imaging modalities in CLIP space, and a small fMRI pool joined to ADNI.

Critical strengths

The geometric observation is correct. Under the linear matching rule, a shifted site contributes the overlap near the anchor, and that subset has a mean far from the site mean. PBMM is a concrete response: keep the matched core, drop sites with an empty core, prefer points near the core, and refuse an outer band that would break the cap. Figure 3 tracks the Lemma 1 surface (1 - alpha) tau. On held-out ABIDE site D1, accuracy is 61.71 for PBMM, 60.70 for matching, and 57.14 for naive pooling. On the six-modality suite, PBMM's deterioration scores stay next to matching (Chest-XRay, OCT17, and RESC all at 4.0) while naive pooling falls to 0 on Chest-XRay. Computing S_k locally and sharing only epsilon_k is a plausible federated split.

Critical weaknesses

Theorem 1 is the opposing monotonicity built into U and V. Objects-as-samples and functors into opposite posets do not add a constraint beyond that one-dimensional picture. With sigma_DP and delta held fixed, epsilon_k is proportional to the mean shift, so the abstract's "over 3x" drop in peak exposure is a drop in that shift. Figures 5 and 15 show band-wise curves; the paper never tables the peak epsilon together with the sigma_DP that was used. Utility is essentially matching's utility. PBMM's ABIDE deterioration score is 11.31 against matching's 12.11. On the heterogeneous suite it is tied or slightly lower on HIS and Liver CT and slightly higher on Brain MRI and RESC. Empty-core sites are removed entirely. There is no comparison with the obvious alternative: match, then keep the most central matched points, with the same budget. DP-SGD on the pooled set and a group-privacy release of the mean are absent. Absolute accuracy stays near 61% on ABIDE. Sections 1 and 3 have enough broken sentences and shifting epsilon notation to impede review on their own.

Questions for authors

See above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.