Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Disjoint Generation of Synthetic Data

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that column-wise partitioning of a tabular dataset, independent generation, and post hoc validated joining improves empirical privacy while keeping utility competitive, especially when different generative models are mixed.

desk verdict Valuable new framework with an honest evaluation, but the headline privacy gain only holds against an adversary who doesn't know the partition; keep the abstract's overclaim in mind. read the letter →

arxiv 2507.19700 v2 pith:TZRP7B6W submitted 2025-07-25 cs.LG

Anton Danholt Lautrup , Muhammad Rajabinasab , Tobias Hyrup , Arthur Zimek , Peter Schneider-Kamp This is my paper · ORCID
classification cs.LG
keywords syntheticdatatabulardisjointgenerativemodelsprivacy-utilitytrade-offempiricalprivacypartitioningmixed-modelgenerationjoiningvalidator
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes generating tabular synthetic data by partitioning the columns into disjoint subsets, training separate generative models on each subset, and reassembling the synthetic partitions post hoc with a learned joining operation that works without shared identifiers. The authors aim to show that this division of labor improves empirical privacy, measured by re-identification risk and membership-inference recall, while the accompanying utility loss can be partially controlled by a validator that scores candidate joins. On benchmark datasets they find that privacy improves as the number of partitions grows, that validated joining recovers utility relative to random concatenation, and that mixing a high-privacy model with a high-utility model yields the best privacy-utility balance. The practical payoff would be that practitioners can compose off-the-shelf generative models to obtain privacy benefits without retraining a single monolithic model or relying on formal differential-privacy guarantees.

What carries the argument

The central object is the joining validator, a binary classifier trained on the real data by pairing authentic rows as positive examples and independently shuffled, mismatched rows as negative examples. At inference time synthetic partitions are randomly concatenated into candidate rows, the validator assigns each candidate a probability of being a valid joint observation, and candidates above an acceptance threshold are removed and added to the output; the remaining pool is reshuffled and the process repeats until the output is large enough. This mechanism carries the argument because it learns which cross-partition combinations are plausible from the training data, converting the privacy gain of fragmentation into a controllable utility trade-off.

What would settle it

Train a DGM with an unoptimized validator (a default random forest without hyperparameter tuning or calibration) on a dataset with strong known cross-partition correlations, then measure Hellinger distance and epsilon-identifiability risk against a random-concatenation baseline; if utility loss equals or exceeds the baseline while privacy gains disappear, the claim that validated joining reconciles privacy and utility fails.

Watch

Extended reading notes

Core claim

The central claim is that column-wise fragmentation of a dataset before synthesis acts as a privacy lever: generating each disjoint partition independently and then reassembling the synthetic pieces makes it harder for an attacker to re-identify real records, at the cost of statistical fidelity. A learned joining validator, trained to distinguish authentic row alignments from randomly shuffled ones, can moderate that cost by admitting only plausible cross-partition combinations. The discovery culminates in mixed-model generation, where a high-privacy model and a high-utility model each handle different partitions, producing synthetic data with downstream accuracy and area-under-the-curve performance close to the high-utility model while empirical re-identification risk approaches that of the privacy-focused model.

Load-bearing premise

The framework's benefits depend on the joining validator being well calibrated and on the partition preserving learnable cross-column correlations; a poorly tuned validator can ruin both marginal and joint fidelity, and an adversary who knows the partition can still compromise privacy.

Editorial extensions

If this is right

  • Increasing the number of disjoint partitions lowers empirical re-identification risk and membership-inference recall, but degrades utility metrics such as correlation-matrix fidelity and holdout AUROC.
  • Using a validated joining procedure instead of random concatenation preserves more utility while keeping some of the privacy benefit.
  • Mixed-model DGMs, pairing a privacy-oriented model with a utility-oriented model, can achieve downstream classification performance close to the utility model while substantially reducing identification risk.
  • Partitioning makes generative models that scale poorly with the number of variables, such as Bayesian-network generators, computationally viable on high-dimensional data.
  • The strength of cross-partition correlation determines which joining strategy works: validated joining helps when partitions share strong dependencies, while concatenation is better when they are nearly independent.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the framework holds up, it suggests that empirical privacy can be treated as a compositional property of data fragmentation rather than only a property of the generative model, which would broaden the design space for privacy-preserving synthesis.
  • The joining validator could potentially be replaced by other alignment or dependency-recovery techniques, such as similarity learning or expectation-maximization, which might recover cross-partition structure even when correlations are weak.
  • A natural stress test beyond the paper is to measure how much privacy actually survives against an adversary who knows the partition structure; the authors state this can compromise privacy, so quantifying that degradation would sharpen the practical guidance.
  • The framework's privacy claims are explicitly empirical, not formal, so a likely next step is composing differentially private submodels and a private validator to obtain end-to-end guarantees without sacrificing the partitioning benefit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Disjoint Generative Models (DGMs) for tabular synthetic data: column-wise partition of the training data, independent generative models per partition, and a post hoc joining step. The joining operation is implemented either by random concatenation or by a learned validator that scores candidate row-level joins, trained on real data with correctly and randomly shuffled row alignments as positive and negative examples. Experiments on seven benchmark datasets plus one high-dimensional dataset report that more partitions improve empirical privacy metrics (ε-identifiability risk, membership inference attack recall) while degrading utility, that the joining validator can partially recover utility, and that mixed-model generation (e.g., synthpop or TabDiff paired with DPGAN) can achieve a middle ground between utility and privacy. The authors also report computational gains for models like Bayesian networks and an appendix study of validator back-ends, threshold settings, calibration, and noise-injection comparisons. The paper explicitly disclaims formal differential privacy guarantees and states that a partition-aware adversary can compromise privacy.

Significance. If the empirical privacy-utility trade-off holds under a realistic threat model, the framework would be practically useful: it lets practitioners combine off-the-shelf generative models without retraining, offers an interpretable knob (the partition assignment and the validator threshold), and is accompanied by reproducible code, repeated experiments across seven datasets, and ablations of validator choices and thresholds. The paper is unusually candid about its limitations, including the absence of end-to-end privacy guarantees and the sensitivity of the validator to calibration and selection bias. However, the central privacy claim is currently stated more strongly than the evaluated threat model supports, and the attribution of the observed privacy gains to the generative process rather than to the validator's selection behavior needs further evidence.

major comments (3)
  1. [Abstract / Section 4.5 (limitation 2)] The headline claim that DGMs "significantly lower the empirical re-identification risk" is stated without the threat-model qualification that the privacy metrics actually assume. The metrics described in Appendix D.2 (ε-identifiability risk and MIA recall) are computed on the final synthetic dataset, under the implicit assumption that the adversary has access to real records but not to the column partition r(·). Limitation 2 in Section 4.5 states that an adversary knowing how data are partitioned "can compromise the privacy of the data easily." Since r(·) is a design input that must be specified to reproduce or use the method, and the implementation is public, a knowledgeable adversary will generally have it. The manuscript needs either (i) a consistent, explicit statement that the privacy claim is against a partition-unaware adversary, or (ii) an experiment with a partition-aware adversary that quantifies the claimed easy compromise (e.g., using the known partition to attempt re-identification from each synthetic partition before joining). As written, the abstract's privacy claim is not supported by the evaluated threat model.
  2. [Appendix C.3 / C.2] The privacy improvements may be an artifact of the joining validator's selection bias rather than a property of disjoint generation itself. Appendix C.3 states that "if the validator model has no notion of authenticity for new, never-before-seen samples, then the apparently improved privacy could stem from a mode collapse introduced during the joining," and Figure 14 shows that stricter acceptance thresholds can improve ε-risk and MIA recall precisely because the admitted set shrinks and overfits to few samples. Appendix C.2 shows that an unoptimized validator can destroy both marginal and joint distribution fidelity. The main-text privacy comparisons (Figures 3, 4, and 7) do not include a control that distinguishes "the generator is more private" from "the validator selected a small, overfit subset." Please add such a control for the headline experiments, for example by measuring ε-risk and MIA recall on an equally sized uniform random subsample of the candidate pool, and by reporting per-dataset distributional fidelity (Hellinger distance, correlation matrix difference) for the admitted sets. Until this is done, the central claim that partitioning and joining cause the privacy gain is not fully established.
  3. [Section 4.5 / Table 3] The claim that mixed-model synthesis "bridges the gap between privacy and utility performance" is supported only by a descriptive comparison on one dataset in Table 3 and by aggregated plots in Figure 7. The paper does not show that the mixed-model configuration lies on or near a Pareto frontier unattainable by any single-model baseline under a privacy constraint (for example, by reporting the best ε-risk each baseline can achieve at a fixed utility level, or by constructing a trade-off curve over validator thresholds and partition schemes). Given the substantial dataset-to-dataset variability noted throughout and the manual choices of partition and validator, a frontier or trade-off analysis would materially strengthen the main claim. As it stands, the evidence shows a favorable middle point, but not that the framework "bridges the gap" in a systematic sense.
minor comments (5)
  1. [Section 4.3] The complexity bound O(n_p (k/n_p)^c) ≤ O(k^c) is stated for "arbitrary positive c," but it holds only for c ≥ 1; for 0 < c < 1 the left-hand side grows as k^c n_p^{1-c}. The factorial bound should be justified or qualified similarly.
  2. [Figures 3–4 vs. Figure 7] The captions are inconsistent about uncertainty: Figures 3 and 4 say standard deviation, Figure 7 says standard error, and Table 3 says "errors of the last significant figure." Please state the uncertainty type uniformly in each caption and table note.
  3. [Appendix B] After Figure 10 the text says "It is difficult to say anything conclusive based on these results," but the following paragraph draws comparative conclusions from the same results. Either soften the conclusions or the caveat.
  4. [Algorithm 1] Algorithm 1 uses a fixed threshold θ, while the experiments use the dynamic threshold scheme described in Appendix A.2 (top-10% initial threshold, adaptive lowering). Please add the dynamic rule to the pseudocode or clearly mark it as a variant, since the pseudocode as written does not reproduce the reported behavior.
  5. [Section 4.5, paragraph on "barely miss the mark"] For (sp,dp), (td,dp), and (ds,dp), the reported ε-risk values of 0.13, 0.12, and 0.15 exceed the 9% reference; calling this "barely miss[ing] the mark" is defensible but should be accompanied by the confidence intervals from Table 3 (e.g., 0.13(2)) so readers can judge the closeness.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the privacy/utility results are empirical and the framework's components are trained on auxiliary tasks rather than on the target metrics.

full rationale

The paper does not fit any parameter to the target evaluation metrics and then report it as a prediction. The joining validator is trained on real data with artificial labels (0 for row-wise shuffled partitions, 1 for authentic joins), which is a legitimate auxiliary component; the final synthetic data are then scored by external metrics such as epsilon-identifiability risk, MIA recall, Hellinger distance, and correlation difference. The reported privacy gain from partitioning is explicitly characterized by the authors as a form of noise: 'This can indeed be viewed as a form of noise introduced in the generative process by fragmenting information' (Section 4.1), and Appendix B compares it to explicit noise injections rather than presenting it as a theorem. The mixed-model results are empirical comparisons against full-data baselines, and the authors do not claim a formal guarantee. Self-citations to SynthEval and to a co-authored PCA metric are present but not load-bearing: the metrics are published, externally usable similarity and utility measures, and the central claims do not reduce to those citations. The authors also state an honest limitation that 'an adversary knowing how the data have been partitioned can compromise the privacy of the data easily' (Section 4.5, limitation 2) and note in Appendix C.3 that apparent privacy gains can stem from mode collapse introduced by the validator; these are robustness caveats, not circular reasoning. No load-bearing step reduces by construction to its own input.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The framework's central claim depends on several domain assumptions about the partition structure, the validator's ability to recover cross-partition dependencies, and the adequacy of empirical privacy metrics. The free parameters are implementation choices (oversampling multiplier, dynamic threshold, and validator hyperparameters) that materially affect the reported results.

free parameters (4)
  • oversampling multiplier = 4
    In A.2, synthetic partitions are oversampled to 4x the training data size to allow enough valid joins before the maximum iteration limit.
  • initial acceptance threshold = top 10%
    In A.2, the dynamic threshold accepts the top 10% of queries in the first iteration; this is an implementation choice that affects composition of the final synthetic dataset.
  • maximum iterations = 100
    In A.2, the validation loop caps at 100 iterations to avoid infinite loops; this impacts how many reshuffling rounds are allowed.
  • validator hyperparameters = optimized per dataset (e.g., 5-20 estimators, depth 10-15)
    In C.2, the random forest validator is hyperparameter-optimized and calibrated; the optimized settings differ from sklearn defaults and materially change the resulting synthetic data fidelity.
assumptions (3)
  • domain assumption Tabular data can be partitioned column-wise without losing the ability to recover inter-column dependencies through the joining operation.
    The framework assumes that generating partitions independently and then joining via a learned validator can approximate the joint distribution sufficiently well. Section 3 and Appendix C.2 show this can fail if the validator is poorly calibrated.
  • domain assumption The validation model is trained on real data with authentic and randomly shuffled joins, and it can generalize to synthetic partitions.
    The validator (Figure 2) learns from real data; the framework assumes that scores on synthetic candidate joins are meaningful. Appendix C.2 discusses overfitting and calibration.
  • domain assumption Empirical privacy metrics (ε-identifiability risk, MIA recall, DCR) are adequate proxies for re-identification risk.
    The paper evaluates empirical, dataset-level privacy rather than formal DP (see 'Please note' in Section 4). The conclusions about privacy improvement rest on these metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disjoint Generation of Synthetic Data." pith.science (2026). https://pith.science/paper/TZRP7B6W

@misc{pith2026250719700,
  author       = {Pith},
  title        = {Pith review of: Disjoint Generation of Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZRP7B6W}},
  note         = {Machine review of arXiv:2507.19700}
}
read the original abstract

We propose a new framework for generating tabular synthetic datasets via disjoint generative models. In this paradigm, a dataset is partitioned into disjoint subsets that are supplied to separate instances of generative models. The results are then combined post hoc by a joining operation that works in the absence of common variables/identifiers. The success of the framework is demonstrated through several case studies and examples on tabular data that help illuminate some of the design choices that one may make. The advantages achieved by the disjoint generation include: i) An observed increase in the empirical measurement of privacy. ii) Increased computational feasibility of certain model types. iii) Ability to generate synthetic data using a mixture of different generative models. Specifically, mixed-model synthesis bridges the gap between privacy and utility performance, providing highly competitive performance on Accuracy and Area Under the Curve for downstream tasks while significantly lowering the empirical re-identification risk.

Figures

Figures reproduced from arXiv: 2507.19700 by the authors.

Figure 1
Figure 1. Disjoint generation conceptual overview. The figure shows how our approach splits training data [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Joining validator training and inference. The validator model is trained in a supervised setting [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Evaluation metrics vs. number of partitions. The figure shows the result of 10 repeated experiments [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Joining operations by metric vs. number of partitions. The figure shows the results of concatenation [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Effect of disjoint generation on running time. The plots show the process time measured for each [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Effect of inter-partition correlation on various metrics. The figure shows results and moving [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Results for mixed-model generation. Top: The result from 10x repeated synthesis of the benchmark [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Baseline comparison with mixed-model generation. This figure shows extra baseline model results: [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Overview of noise injection comparisons. The figure presents a conceptual overview of the three [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Evaluation metrics vs. noise degree. The figure shows high level results of noise injection (average [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Effect of different validators. This figure shows the effect of using different validator models on [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Calibration plots showing optimisation in effect. The three calibration displays show the difference [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Impact of validation on distributional similarity. The figures demonstrate how using validation [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Evaluation metrics of datasets built at different validator thresholds. The line graphs show the [PITH_FULL_IMAGE:figures/full_fig_p025_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages

  1. [1]

    A.1 Python library For investigating the potential of DGMs, and exploring various aspects, we implemented a basic library for disjoint synthetic data generation that interfaces various generative model frameworks, namely, Synthc- ity (Qian et al., 2023), DataSynthesizer (Ping et al., 2017), TabDiff (Shi et al., 2025), and synthpop (Nowok et al., 2016), wi...

    work page 2023

  2. [6]

    European Medicines Agency

    doi: 10.1093/jamiaopen/ooac083. European Medicines Agency. External guidance on the implementation of the Eu- ropean Medicines Agency policy on the publication of clinical data for medici- nal products for human use.https://www.ema.europa.eu/en/human-regulatory/ marketing-authorisation/clinical-data-publication/support-industry/ external-guidance-implemen...

    work page doi:10.1093/jamiaopen/ooac083

  3. [8]

    doi: 10.1007/978-3-031-09342-5_17

    Springer. doi: 10.1007/978-3-031-09342-5_17. Joao Fonseca and Fernando Bacao. Tabular and latent space synthetic data generation: a literature review. J. Big Data, 10(1):115, July

    work page doi:10.1007/978-3-031-09342-5_17

  4. [9]

    Lasse Hansen, Nabeel Seedat, Mihaela van der Schaar, and Andrija Petrovic

    doi: 10.1186/s40537-023-00792-7. Lasse Hansen, Nabeel Seedat, Mihaela van der Schaar, and Andrija Petrovic. Reimagining synthetic tabular data generation through data-centric AI: A comprehensive benchmark. InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA,...

    doi:10.1186/s40537-023-00792-7

  5. [10]

    A simple and practical algorithm for differentially private data release

    15 Published in Transactions on Machine Learning Research (06/2026) Moritz Hardt, Katrina Ligett, and Frank McSherry. A simple and practical algorithm for differentially private data release. InAdvances in Neural Information Processing Systems 25, NIPS 2012, volume 25, pp. 2348–2356, Red Hook, NY, USA,

    work page 2026

  6. [11]

    2022.04.053

    doi: 10.1016/j.neucom. 2022.04.053. Charles A.R. Hoare. Algorithm 64: Quicksort.Commun. ACM, 4(7):321, July

    doi:10.1016/j.neucom 2022

  7. [14]

    hepatitis

    There are several noteworthy behaviours to con- sider: The plainscikit-learnmodel and the suboptimal model exhibit little change before the threshold reaches approximately halfway through the bell curves in their probability spectrum. On the other side, the performance on most utility metrics suddenly deteriorates, with some improvement to privacy, indica...

    work page 2025

  8. [15]

    Anton Danholt Lautrup, Tobias Hyrup, Arthur Zimek, and Peter Schneider-Kamp

    PMLR. Anton Danholt Lautrup, Tobias Hyrup, Arthur Zimek, and Peter Schneider-Kamp. Systematic review of generative modelling tools and utility metrics for fully synthetic tabular data.ACM Comput. Surv., 57 (4):90:1–90:38, December 2024a. doi: 10.1145/3704437. Anton Danholt Lautrup, Tobias Hyrup, Arthur Zimek, and Peter Schneider-Kamp. Syntheval: a framewo...

    doi:10.1145/3704437

Show all 41 references
  1. [17]

    Beata Nowok, Gillian M

    doi: 10.2196/medinform.7744. Beata Nowok, Gillian M. Raab, and Chris Dibben. synthpop: Bespoke creation of synthetic data in R.J. Stat. Softw., 74(11):1–26,

    doi

  2. [20]

    Preprint athttps://arxiv.org/abs/2301. 07573. 16 Published in Transactions on Machine Learning Research (06/2026) Muhammad Rajabinasab, Anton Danholt Lautrup, and Arthur Zimek. Metrics for inter-dataset similarity with example applications in synthetic data and feature selecti...

    2026

  3. [21]

    doi: 10.1137/1.9781611978520.57

    SIAM. doi: 10.1137/1.9781611978520.57. Debbie Rankin, Michaela Black, Raymond Bond, Jonathan Wallace, Maurice Mulvenna, and Gorka Epelde. Reliability of supervised machine learning using synthetic data in health care: Model to preserve privacy for data sharing.Med. Inform., 8(...

    doi

  4. [22]

    Vicenç Torra I

    doi: 10.2196/18910. Vicenç Torra I. Reventós. OWA operators in data modeling and reidentification.IEEE Trans. Fuzzy Syst., 12(5):652–660,

    doi

  5. [25]

    Juntong Shi, Minkai Xu, Harper Hua, Hengrui Zhang, Stefano Ermon, and Jure Leskovec

    doi: 10.3389/frai.2022.918813. Juntong Shi, Minkai Xu, Harper Hua, Hengrui Zhang, Stefano Ermon, and Jure Leskovec. Tabdiff: a mixed-type diffusion model for tabular data generation. InProceedings of The Thirteenth International Conference on Learning Representations, ICLR 202...

    2022

  6. [27]

    Sandhya Tripathi, Bradley A

    Preprint athttps://arxiv.org/abs/2501.12012. Sandhya Tripathi, Bradley A. Fritz, Mohamed Abdelhack, Michael S. Avidan, Yixin Chen, and Christo- pher Ryan King. Multi-view representation learning for tabular data integration using inter-feature rela- tionships.J. Biomed. Inform...

    arXiv

  7. [28]

    Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Mihaela van der Schaar

    doi: 10.1016/j.jbi.2024.104602. Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Mihaela van der Schaar. DECAF: generating fair synthetic data using causally-aware generative networks. InProceedings of the 35th International Conference on Neural Information Processing Sy...

    2024

  8. [30]

    PMLR. D. R. Wilson. Beyond probabilistic record linkage: Using neural networks and complex features to improve genealogical record linkage. InThe 2011 International Joint Conference on Neural Networks, IJCNN 2011, pp. 9–14, New York, NY, USA,

    2011

  9. [32]

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni

    Preprint athttps://arxiv.org/abs/1802.06739. Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Modeling tabular data using conditional GAN. InAdvances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Sy...

    2019 arXiv

  10. [33]

    17 Published in Transactions on Machine Learning Research (06/2026) Andrew Yale, Saloni Dash, Ritik Dutta, Isabelle Guyon, Adrien Pavao, and Kristin P. Bennett. Generation and evaluation of privacy preserving synthetic health data.Neurocomputing, 416:244–255, November

    2026

  11. [34]

    Jinsung Yoon, Lydia N

    doi: 10.1016/j.neucom.2019.12.136. Jinsung Yoon, Lydia N. Drumright, and Mihaela van der Schaar. Anonymization through data synthesis using generative adversarial networks (ADS-GAN).IEEE J. Biomed. Health Informatics, 24(8):2378–2388,

    2019 doi

  12. [35]

    Xun Yuan, Yang Yang, Prosanta Gope, Aryan Pasikhani, and Biplab Sikdar

    doi: 10.1109/jbhi.2020.2980262. Xun Yuan, Yang Yang, Prosanta Gope, Aryan Pasikhani, and Biplab Sikdar. Vflgan: Vertical federated learning-based generative adversarial network for vertically partitioned data publication.Proc. Priv. Enh. Technol., 2024(4):840–858, October

    2020

  13. [36]

    A Implementation details This appendix contains details on the implementation of the disjoint generation framework that we made and use for the experiments6

    doi: 10.56553/popets-2024-0144. A Implementation details This appendix contains details on the implementation of the disjoint generation framework that we made and use for the experiments6. Some practical choices made in the implementation are left out of the main text since t...

    2024 doi

  14. [38]

    18 Published in Transactions on Machine Learning Research (06/2026) using the dynamic and tied-down behaviours to be insignificant and minor (≲5%), in favour of the dynamic case

    holds the implementation, tutorial, and codebooks to reproduce the experimental results. 18 Published in Transactions on Machine Learning Research (06/2026) using the dynamic and tied-down behaviours to be insignificant and minor (≲5%), in favour of the dynamic case. Because t...

    2026

  15. [39]

    φ1,φ 2,φ 3

    Columnwise shuffling entails selectingφ1·nof the rows, and then shuffle their values independently and column-wise. Random replacement replacesφ2·nof the column values with the column mode, and Gaussian noise involves replacing the cell valuesxwithx′ =x+N(0,φ 3σx). In Figure 1...

    2026

  16. [40]

    moving parts

    and TabDDPM (Kotelnikov et al., 2023), which natively incorporate noise, compression, and denoising within a coherent end-to-end methodology. These models provide a principled reference for understanding how information degrada- tion and recovery are handled in mature generati...

    2023

  17. [1961]

    doi: 10.1145/366622. 366644. Tobias Hyrup, Anton Danholt Lautrup, Arthur Zimek, and Peter Schneider-Kamp. A systematic review of privacy-preserving techniques for synthetic tabular health data.Discover Data, 3(1), March

    doi

  18. [1962]

    Tabddpm: Modelling tabular data with diffusion models

    Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. Tabddpm: Modelling tabular data with diffusion models. InInternational Conference on Machine Learning, ICML 2023, volume 202 ofProc. Mach. Learn. Res., pp. 17564–17579, Honolulu, Hawaii, USA,

    2023

  19. [1965]

    doi: 10.1090/s0025-5718-1965-0178586-1. Fida K. Dankar, Mahmoud K. Ibrahim, and Leila Ismail. A multi-dimensional evaluation of synthetic data generators.IEEE Access, 10:11147–11158,

    1965 doi

  20. [2000]

    Junghye Lee, Jimeng Sun, Fei Wang, Shuang Wang, Chi-Hyuck Jun, and Xiaoqian Jiang

    doi: 10.1007/978-1-4612-1166-2. Junghye Lee, Jimeng Sun, Fei Wang, Shuang Wang, Chi-Hyuck Jun, and Xiaoqian Jiang. Privacy-preserving patient similarity learning in a federated environment: Development and analysis.JMIR Med. Inform., 6 (2):e20, April

    doi

  21. [2004]

    Peter Schneider-Kamp, Anton Danholt Lautrup, and Tobias Hyrup

    doi: 10.1109/tfuzz.2004.834814. Peter Schneider-Kamp, Anton Danholt Lautrup, and Tobias Hyrup. Synthesizers: A meta-framework for generating and evaluating high-fidelity tabular synthetic data. InProceedings of the 19th International Conference on Software Technologies, ICSOFT...

    2004

  22. [2006]

    Khaled El Emam, Lucy Mosquera, and Xi Fang

    doi: 10.1007/11681878_14. Khaled El Emam, Lucy Mosquera, and Xi Fang. Validating a membership disclosure metric for synthetic health data.JAMIA Open, 5(4):ooac083, 10

    doi

  23. [2011]

    doi: 10.1109/ijcnn.2011.6033192

    IEEE. doi: 10.1109/ijcnn.2011.6033192. Liyang Xie, Kaixiang Lin, Shu Wang, Fei Wang, and Jiayu Zhou. Differentially private generative adversarial network,

    2011

  24. [2016]

    Haoyue Ping, Julia Stoyanovich, and Bill Howe

    doi: 10.18637/jss.v074.i11. Haoyue Ping, Julia Stoyanovich, and Bill Howe. Datasynthesizer: Privacy-preserving synthetic datasets. InProceedings of the 29th International Conference on Scientific and Statistical Database Management, SSDBM ’17, pp. 42:1–42:5, New York, NY, USA,

    doi

  25. [2017]

    doi: 10.1145/3085504.3091117

    ACM. doi: 10.1145/3085504.3091117. Zhaozhi Qian, Bogdan-Constantin Cebere, and Mihaela van der Schaar. Synthcity: facilitating innovative use cases of synthetic data in different data modalities,

  26. [2018]

    DP-CTGAN: differentially private medical data generation using ctgans

    Mei Ling Fang, Devendra Singh Dhami, and Kristian Kersting. DP-CTGAN: differentially private medical data generation using ctgans. InArtificial Intelligence in Medicine, AIME 2022, volume 13263 ofLect. Notes Comput. Sci., pp. 178–188, Cham,

    2022

  27. [2019]

    Paul Tiwald, Ivona Krchova, Andrey Sidorenko, Mariana Vargas Vieyra, Mario Scriminaci, and Michael Platzer

    doi: 10.1111/insr.12353. Paul Tiwald, Ivona Krchova, Andrey Sidorenko, Mariana Vargas Vieyra, Mario Scriminaci, and Michael Platzer. Tabularargn: A flexible and efficient auto-regressive framework for generating high-fidelity syn- thetic data,

    doi

  28. [2020]

    doi: 10.1145/3411170.3411243

    ACM. doi: 10.1145/3411170.3411243. Shaoming Duan, Chuanyi Liu, Peiyi Han, Xiaopeng Jin, Xinyi Zhang, Tianyu He, Hezhong Pan, and Xiayu Xiang. Ht-fed-gan: Federated generative model for decentralized tabular data synthesis.Entropy, 25(1): 88, December

  29. [2021]

    Curran Associates Inc. David S. Watson, Kristin Blesch, Jan Kapar, and Marvin N. Wright. Adversarial random forests for density estimation and generative modeling. InInternational Conference on Artificial Intelligence and Statistics, AISTATS 2023, Proc. Mach. Learn. Res., pp. ...

    2023

  30. [2022]

    Irina Deeva, Petr D

    doi: 10.1109/access.2022.3144765. Irina Deeva, Petr D. Andriushchenko, Anna V. Kalyuzhnaya, and Alexander V. Boukhanovsky. Bayesian networks-based personal data synthesis. InProceedings of the 6th EAI International Conference on Smart Objects and Technologies for Social Good, ...

    2022

  31. [2023]

    Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam D

    doi: 10.3390/e25010088. Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam D. Smith. Calibrating noise to sensitivity in private data analysis. InTheory of Cryptography, Third Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March 4-7, 2006, Proceedings, volu...

    2006 doi

  32. [2024]

    doi: 10.5220/0012856000003753

    SciTePress. doi: 10.5220/0012856000003753. Jingpu Shi, Dong Wang, Gino Tesei, and Beau Norgeot. Generating high-fidelity privacy-conscious synthetic patient data for causal effect estimation with multiple treatments.Frontiers Artif. Intell., 5,

    doi

  33. [2025]

    Xiangjian Jiang, Nikola Simidjievski, and Mateja Jamnik

    doi: 10.1007/s44248-025-00022-w. Xiangjian Jiang, Nikola Simidjievski, and Mateja Jamnik. Tabstruct: Measuring structural fidelity of tabular data. InThe Fourteenth International Conference on Learning Representations, ICLR 2026,

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.