acceptodds
Under review as a conference paper at ICLR 2027

Differentially Private Vine Copulas for Preserving Higher-Order Dependence Structures

Abstract

Generating high-utility synthetic data under differential privacy (DP) remains challenging when complex dependence and higher-order interactions are present. Traditional marginal and graph-based DP methods can preserve low-dimensional structure, such as univariate marginals and pairwise correlations, but often fail to recover higher-order interactions. While DP deep generative models such as GANs can, in principle, capture complex dependence, they are often unstable and difficult to tune. We propose Differentially Private Vine Copulas (DPRIVVE), an end-to-end differentially private synthetic data generator that models higher-order dependence via conditional propagation. DPRIVVE combines differentially private marginal transformations to construct DP pseudo-observations without data leakage, private D-vine ordering, and private copula family and parameter estimation. Notably, DPRIVVE more accurately captures higher-order interactions relative to competing DP baselines. Moreover, these improvements consistently hold over varying privacy budgets and dependence structures while preserving standard utility metrics.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
differential privacy, synthetic data generation, vine copulas, higher-order data dependence, interpretability

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

This paper propose Differentially Private Vine Copulas (DPRIVVE), an end-to-end differentially private synthetic data generator that models higher-order dependence via conditional propagation. This paper argues that, high-order dependences among data features, often being low-signal effects, can easily be wiped out via standard DP mechanisms. Therefore, the paper introduces vine copulas, a previously studied approach in other fields, to provide a framework for multivariate modeling by decomposing a joint distribution into univariate marginals and a sequence of bivariate copulas arranged in a vine structure. Experiment on both real and controlled higher-order simulation settings, showing the proposed DPRIVVE method's capability in modeling high-order dependencies among data features.

Critical strengths
  1. I like the controlled higher-order simulation datasets in the experimental part that directly demonstrates the effectiveness of the proposed method in the problem it aims to solve: modeling the high-order dependences among different sample features.
  2. The paper is clearly written and easy to follow. The way it introduces vine copulas is easy to follow and quickly gives me a high level sense of what the this part is for.
  3. The assumptions that are based on the mathematical formulation part is clearly stated and reasonable.
Critical weaknesses
  1. Although I like the controlled higher-order simulation datasets in the experimental part, they are way too easy and seem more to me like a toy demo rather than something that should appear in the main experiments. Other datasets included in the main experiments are also simple datasets (often tabular datasets) with limited number of features to deal with. Therefore, I think the experimental part is too weak given this interesting question. I would like to see experiments on (1) harder real datasets with more features compared to the current ones; (2) even image datasets that may also benefit from the high-order dependences within the dataset.
  2. Like show in Figure 5 and Figure 14, the results lines are flat with the change of epsilon, which does not reflect how larger noise under stricter noise will affect the method's performance. So I would like to see results with smaller ε\varepsilon values that should bring some more changes to the lines.
  3. How is (FjDP^)−1(\hat{F_j^{DP}})^{-1} calculated is not described in the paper.
  4. The models used in the experimental part is too easy, like decision trees. I do not doubt the importance of simple models, but I would like to know if this method has a potential to be still effective in larger and more complex deep learning models, even transformers models. Or, will DPRIVVE's effect be wiped out if applied with more complexed models?
Questions for authors
  1. Why don't you put more results to the main paper? currently it is shorter than 9 pages, but you still have a lot more results in the appendix.
Reviewer_X8TiOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces DPRIVVE, a differentially private synthetic data generation framework built on D-vine copulas. By integrating private marginal estimation, vine structure ordering, and copula fitting, the method effectively preserves higher-order data dependencies. Extensive experiments across synthetic and real-world datasets demonstrate that DPRIVVE achieves superior interaction recovery compared to existing differential privacy (DP) generators.

Critical strengths
  1. Preserving higher-order dependencies under differential privacy is an important and practically meaningful challenge.

  2. The vine-copula formulation provides a structured alternative to neural generative models.

Critical weaknesses
  1. The proof treats pseudo-observation construction as post-processing, although (Uij=F^jDP(xij))(U_{ij}=\hat F_j^{DP}(x_{ij})) still depends on the original private records. The claimed Kendall's tau sensitivity also requires clarification under add/remove adjacency. Consequently, the end-to-end privacy guarantee is not rigorously established as written.

  2. The exponential mechanism uses clipped copula log-likelihoods, but the paper does not explain how candidate copula parameters are obtained before family selection. If these parameters are estimated nonprivately, the claimed sensitivity bound of (2L) does not immediately apply.

Questions for authors
  1. How is constructing (Uij=F^jDP(xij))(U_{ij}=\hat F_j^{DP}(x_{ij})) justified as post-processing when it accesses the original records? Can the authors provide a formal adaptive-composition argument?

  2. Why is AIM restricted to maximum marginal degree three when the main continuous simulation evaluates dependence involving four variables? Would allowing degree-four marginals change the conclusions?

  3. How are the copula parameters obtained when computing the family-selection log-likelihoods?

Reviewer_5G2xOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates differentially private (DP) synthetic data generation. In particular, it proposes using vine copulas to model higher-order dependencies among different attributes. As I understand it, the method first privatizes the marginal distributions by adding Laplace noise to histogram counts, from which differentially private cumulative distribution functions (CDFs) are constructed. These privatized CDFs are then applied to the original data to obtain pseudo-observations. Next, the method computes differentially private estimates of Kendall's τ on these pseudo-observations to measure pairwise dependencies and determine the variable ordering in the D-vine structure. Given this ordering, the method fits pair copulas layer by layer, using the exponential mechanism to privately select copula families and privatized Kendall's τ estimates to determine their parameters. Higher-level conditional dependencies are modeled through h-functions, which recursively construct conditional pseudo-observations for subsequent vine levels.

Critical strengths

As this submission primarily focuses on proposing a new technical method, I would place greater emphasis on its methodological novelty and the insights it provides. To the best of my knowledge, the proposed use of D-vine copulas for differentially private synthetic data generation introduces some novel methodological components. The paper also reports empirical performance improvements over several baselines, which is encouraging.

Critical weaknesses

As this is a privacy paper, I place particular emphasis on the clarity of the presentation and the rigor of the privacy proof. However, in these respects, the paper appears somewhat premature and is difficult to follow, particularly in the method section. Moreover, the privacy proof is relatively vague and not sufficiently self-contained, making it difficult to assess the validity of the claimed privacy guarantees. I am also concerned that the methodological contribution is somewhat incremental and that the experimental section offers limited additional insights. Taken together, these concerns would substantially lower my overall rating.

Questions for authors
  • What are the key differences between the proposed method and existing DP copula-based mechanisms? What specific advantages does the proposed approach offer? Additional ablation studies would also help clarify the contributions of individual components.

  • The experimental evaluation appears somewhat limited in its coverage of existing DP synthetic data generation methods. Several relevant approaches are missing, to name a few: RAP (ICML 2021), RAP++ (NeurIPS 2022), PrivSyn (USENIX Security 2021), PrivMRF (VLDB 2021), Private-GSD (ICML 2023), and Tab-PE (ICML 2026). Beyond this, a deeper investigation into the practical necessity of modeling higher-order dependencies would be valuable. Many existing DP synthetic data generation methods focus on preserving low-dimensional dependencies, which can already provide good utility for various downstream tasks. In contrast, the proposed method introduces additional privacy-budget costs and methodological complexity to model higher-order dependencies. Under what conditions do these additional costs lead to meaningful improvements in practical utility? A more comprehensive comparison with recent methods, together with an analysis of the privacy–utility–complexity trade-off, would help better justify the proposed D-vine approach.

RAP: Sergul Aydore, William Brown, Michael Kearns, Krishnaram Kenthapadi, Luca Melis, Aaron Roth, and Ankit A. Siva. "Differentially Private Query Release Through Adaptive Projection." 38th International Conference on Machine Learning (ICML), 2021. https://proceedings.mlr.press/v139/aydore21a.html

RAP++: Giuseppe Vietri, Cedric Archambeau, Sergul Aydore, William Brown, Michael J. Kearns, Aaron Roth, Ankit Siva, Shuai Tang, and Steven Z. Wu. "Private Synthetic Data for Multitask Learning and Marginal Queries." 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. https://papers.neurips.cc/paper_files/paper/2022/hash/7428310c0f97f1c6bb2ef1be99c1ec2a-Abstract-Conference.html

PrivSyn: Zhikun Zhang, Tianhao Wang, Ninghui Li, Jean Honorio, Michael Backes, Shibo He, Jiming Chen, and Yang Zhang. "PrivSyn: Differentially Private Data Synthesis." 30th USENIX Security Symposium (USENIX Security), 2021. https://www.usenix.org/conference/usenixsecurity21/presentation/zhang-zhikun

PrivMRF: Kuntai Cai, Xiaoyu Lei, Jianxin Wei, and Xiaokui Xiao. "Data Synthesis via Differentially Private Markov Random Fields." Proceedings of the VLDB Endowment (PVLDB), 14(11):2190–2202, 2021. https://www.vldb.org/pvldb/vol14/p2190-cai.pdf

Private-GSD: Terrance Liu, Jingwu Tang, Giuseppe Vietri, and Zhiwei Steven Wu. "Generating Private Synthetic Data with Genetic Algorithms." 40th International Conference on Machine Learning (ICML), 2023. https://proceedings.mlr.press/v202/liu23ag.html

Tab-PE: Toan Tran, Arturs Backurs, Zinan Lin, Victor Reis, Li Xiong, and Sergey Yekhanin. "Differentially Private Synthetic Data via APIs 4: Tabular Data." 43rd International Conference on Machine Learning (ICML), 2026. https://proceedings.mlr.press/v306/tran26a.html

  • The proposed method involves several hyperparameters and design choices, including privacy-budget allocation, marginal discretization, and vine truncation levels. How sensitive is the performance to these choices, and how well do the recommended settings generalize across different datasets and privacy budgets?

  • The privacy proof is somewhat unclear, especially for the higher levels of the D-vine. The method uses h-functions to transform the data before computing Kendall's τ\tau at each level. Is the claimed sensitivity bound of (4/(n+1)) still valid after these transformations, particularly when they depend on previously estimated parameters? Moreover, the paper appears to treat conditional propagation through h-functions as post-processing, even though these computations may still involve the original sensitive data. Could the authors provide a more rigorous justification of the sensitivity bound and clarify how the privacy costs of these operations are accounted for, potentially through adaptive composition?

  • Minor Point: Figure 1 contains clear AI-generated visual artifacts that should be corrected.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.