acceptodds
Under review as a conference paper at ICLR 2027

Differentially Private Wasserstein Barycenters

Abstract

The Wasserstein barycenter is defined as the mean of a set of probability measures under the optimal transport metric, and has numerous applications in machine learning. In practice these input measures are empirical distributions, often built from sensitive datasets. We present, to our knowledge, the first algorithms for computing Wasserstein barycenters under differential privacy. Empirically, on large-scale US population datasets, our methods produce high-quality private barycenters with strong accuracy-privacy tradeoffs. As a byproduct of our analysis we give a simple reduction that can significantly accelerate the runtime for non-private barycenters, which may be of broader interest.

Primary area:
optimization
Keywords:
differential privacy, optimal transport, Wasserstein barycenters, clustering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RyQCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies DP Wasserstein barycenters of multiple empirical distributions, with a prescribed number of equally weighted output atoms. It proposes two approaches: constructing DP Wasserstein coresets and learning shared private cluster centers with noisy group-specific counts. Both use a non-DP barycenter solver after preprocessing. The theoretical contributions include approximation guarantees conditional on the underlying solver, an excess-cost lower bound through private mean estimation, and a clustering reduction for accelerating non-private computation. Experiments on census-derived spatial data and higher-dimensional synthetic data evaluate objective values, runtime, and privacy–utility tradeoffs. The reported results show some advantages of proposed method.

Critical strengths
  1. A useful, modular reduction from private clustering to private barycenter computation.

  2. Empirical evidence of computational and privacy-utility benefits

Critical weaknesses
  1. The novelty claim is too broad, and the closest prior work is insufficiently addressed.

The claim of providing the first differentially private Wasserstein barycenter algorithms needs qualification in light of earlier private barycenter methods, including Xian et al. (2024) and Han et al. (2024). The paper needs a sharper distinction between its general multidimensional, record-level privacy setting and existing results.

  1. Important gaps remain between the stated guarantees and the algorithmic specification.

Proposition 3 is not formulated appropriately for arbitrary approximate solver outputs, and duplicating atoms to represent rational weights can cause exponential growth in input size. In addition, the experimental description does not sufficiently specify the private clustering implementation and bounded-domain preprocessing.

Questions for authors

Except the weaknesses,

  1. Which private clustering implementation and privacy-budget allocation were used in the experiments?

  2. What normalization or clipping rules were applied to the geographic coordinates and Gaussian samples to satisfy Assumption 1?

  3. Does “exact barycenter” refer to exact transport subproblems or global barycenter optimality, and are all reported objective values evaluated using unregularized transport?

  4. Do the reported runtimes include clustering and preprocessing, and are initialization and stopping criteria shared across the compared methods?

Reviewer_r5Q5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

Main idea : to provide Differentially Private algorithms for the Wasserstein barycenter problem and establish properties and guaranties of such algorithms.

The paper is structured in the following way :

  • the introduction mostly convers the related works and states the contributions of the article
  • the preliminary part is very small : it defines the Wasserstein distance and gives a definition of an differentially private algorithm.
  • matrix Pi is used for Projection and transport plan*
  • the next section define the Wassertsein barycenter problem and gives assumption on the support size and shape of the barycenter measure, as well as defining Neighbouring datasets.

It is clear that this two parts could have been regrouped and restructured (for example OT subsection, DP subsection)

  • the next section is a first proposition of a DP algorith. As the Algorithm 1 isn’t refered in the text it is not exactly clear from a reader point-of-view. From what I understand, each measure is given a private proxy that is projected on a subspace. From these measures an approximate Wasserstein barycenter is computed. The barycenter in the original space is then reconstructed from the projected barycenter and the original measure following a reference (Izzo et al.).

The algorithm 1, is mostly an adaptation of Izzo et al. with private coreset. In the text there is also a reference to the algorithm of He et al., but this article contains several algorithms. From what I understand, the author wants to adapt the algorithm from Izzo et al., and need a DP algorithm to obtain the proxy. This task is done in He et al. 2023, with guarantees. Then the privacy is propagated through parallel composition. Proposition 1 DP-caracterizes the proxy operations. Theorem 1 the whole algorithm.

  • Section 5 introduces a DP clustering step and numerous steps of privacy to construct the proxy measures (in the same dimensionnal space). The authors then provides a Theorem stating the privacy guarantee of the algorithm.

  • Section 6 presents experiments on the US Census data and one case of "highly" dimensional (d=25) data.

  • There are no conclusion or perpectives.

Critical strengths

The proposal of a guaranteed DP WB is a good idea, and the theoritical work appears sound.

However the paper needs works as its actual structure and writing lacks clarity.

Critical weaknesses

My opinion is that the article is "not yet finished" and should be majorly restuctured to enounce more clearly the contributions.

For example, Algorithm 1 is not referred in the text nor is the fact that Algorithm 1 adapts to DP an algorithm from Izzo et al.. Algorithm 1 is the core of the first contribution.

In a similar way, if the 'non-private' clustering WB algorithm provides an approximation of the WB and is new it should be presented and studied either in the article or the appendix and then DP modification could be presented and analysed.

Moreover the experimental part lacks structure and insights. There a lot of figures with the US census data, which is good example, but in order to discuss the theoritical part, synthetic data could be used to produce more experiments than the one provided. A classical way is to introduce "questions" and build the experiments to give answers.

It is a good practice to offer the reader a short summary of each section of the paper at the end of the inroduction and to close the paper with a very short summary of the whole article and perspectives. Another good pratice is to give sketch of the proof for some of the main results.

Questions for authors

One of the hidden assumptions of the work is that all measure share the supprt cardinality (n). Could this assumption be removed ?

Does the clustering approach converges exactly to WB when the number of cluster tends to kn ? Could you provide names and insights to the different parts (3-6, 7-11) of Algorithm 2 ?

Reviewer_9LpdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the problem of computing Wasserstein barycenters under differential privacy. The authors propose two approaches: a private coreset-based method that constructs private approximations of input distributions, and a private clustering-based method that compresses the data using differentially private clustering and noisy histograms before applying a standard barycenter solver. They establish theoretical utility guarantees for both approaches, derive a lower bound on the privacy-induced error, and provide a clustering-based reduction that can also accelerate non-private barycenter computation. Experiments on US census and synthetic datasets demonstrate the effectiveness of the proposed methods and illustrate their privacy–utility tradeoffs.

Critical strengths
  1. The paper develops two differentially private algorithms for Wasserstein barycenters with provable utility guarantees. In particular, the clustering-based reduction is a useful contribution, as it enables existing non-private barycenter solvers to operate on privately compressed data while retaining theoretical guarantees.
  2. Beyond the algorithmic upper bounds, the paper establishes a lower bound on privacy-induced error and provides a reduction that can accelerate non-private barycenter computation. These complementary results provide a more complete understanding of the problem and broaden the applicability of the proposed framework.
Critical weaknesses
  1. The paper does not discuss Differentially Private Post-Processing for Fair Regression (Xian et al., ICML 2024), which also computes Wasserstein barycenters from privately estimated distributions. Although the problem settings and theoretical guarantees differ, the underlying private-preprocessing framework is closely related. A more explicit comparison would help clarify the novelty and distinct contributions of the proposed methods.
  2. The experimental implementation of the private clustering algorithm is not clearly described, making it difficult to assess its consistency with the theoretical guarantees. Furthermore, the lack of comparisons with alternative private baselines and end-to-end runtime measurements limits the evaluation of the proposed methods' practical effectiveness.
Questions for authors
  1. The relationship to Differentially Private Post-Processing for Fair Regression (Xian et al., ICML 2024) deserves further discussion. How do the proposed methods differ in their algorithmic design and theoretical guarantees, and what are the main additional technical contributions?
  2. It would be helpful to clarify which private clustering algorithm is used in the experiments and whether its implementation satisfies the assumptions underlying Theorem 2.
  3. The experiments demonstrate computational benefits, but the end-to-end efficiency of the private methods remains unclear. Could the authors clarify the overall computational cost and how it compares with alternative private approaches under comparable privacy budgets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.