Privacy-Preserving Tabular Synthetic Data Generation Using TabularARGN

Andrey Sidorenko; Paul Tiwald

REVIEW 4 major objections 4 minor 57 references

Privacy-Preserving Tabular Synthetic Data Generation Using TabularARGN

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read TabularARGN, a discretization-based auto-regressive generative network, produces synthetic mixed-type tabular data whose statistical fidelity, machine-learning utility, and resistance to synthetic-vs-real detection match GAN- and diffusion-

desk verdict A coherent abstract and positioning for an autoregressive tabular generator, but the manuscript we have stops before any method or evidence, so the central fidelity claim is unverifiable. read the letter →

arxiv 2508.06647 v1 pith:YI3UMZNK submitted 2025-08-08 cs.LG

Andrey Sidorenko , Paul Tiwald This is my paper

classification cs.LG

keywords syntheticdatatabularauto-regressivegenerativenetworkdiscretizationmembershipinferenceattackprivacy-utilitytrade-offdensityestimationmixed-type

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

The paper introduces TabularARGN, a neural network that generates synthetic tabular data by factorizing the joint distribution of a table into per-column conditional probabilities over discretized value ranges. The authors aim to show that this auto-regressive, discretization-based design can match the data quality of GAN- and diffusion-based generators while being substantially cheaper to train. They also argue that privacy in synthetic data should be audited with membership-inference attacks rather than distance-to-closest-record metrics, and that TabularARGN fares well under such an audit. If these claims hold, organizations needing to share sensitive tables could use a simpler model family without sacrificing fidelity or a defensible privacy profile.

What carries the argument

The central object is TabularARGN (Tabular Auto-Regressive Generative Network), an auto-regressive neural density estimator that factors the joint distribution $P(x_1,\dots,x_d)$ into a product of conditional probabilities $\prod_i P(x_i \mid x_{<i})$. Discretization maps each column's values into a finite set of bins, so generation becomes sequential sampling: draw the first column, then condition on it for the next, and so on. This converts mixed-type numeric and categorical columns into a common categorical representation, making each conditional step a classification problem and letting one network handle heterogeneous tables without per-column loss engineering.

What would settle it

Run a preregistered benchmark on a public mixed-type tabular dataset using a fixed binning scheme and column order, compare TabularARGN against CTGAN and TabDDPM on downstream classifier utility, and run a shadow-model membership-inference attack at a 1% false-positive target. If TabularARGN trails the baselines on utility by a meaningful margin, or if the attack's true-positive rate at 1% false-positive rate exceeds random-chance expectation, the paper's competitive-fidelity and privacy-robustness claims fail together.

Watch

Extended reading notes

Core claim

TabularARGN treats tabular synthetic-data generation as a density-estimation problem and approximates the joint distribution of a table as a sequence of conditional probabilities, one column at a time, using explicit discretization of value ranges instead of treating rows as text. The paper's central claim is that this design yields synthetic data whose statistical similarity, downstream machine-learning utility, and resistance to detection are competitive with current GAN- and diffusion-based tabular generators, while requiring far less computation. Its second claim is that privacy should be measured by systematic membership-inference attacks, and that under such attacks TabularARGN-generat

Load-bearing premise

The paper's conclusions depend on the premise that modeling a table's columns one at a time, each over discretized value ranges, can capture the joint distribution accurately enough to match GAN and diffusion baselines, and that membership-inference attacks are the right yardstick for disclosure risk.

Editorial extensions

If this is right

Synthetic tabular data with fidelity comparable to GAN and diffusion baselines could be produced without adversarial training or diffusion sampling, lowering the compute barrier for large synthetic tables.
Privacy audits of synthetic-data generators can shift away from nearest-record distance metrics toward membership-inference attacks, which the paper argues are more informative.
A single architecture can model mixed-type tables by discretizing every column, avoiding bespoke loss functions for numeric versus categorical columns.
Sequential sampling makes conditional generation natural: drawing synthetic rows conditioned on a fixed column value is a direct by-product of the generation loop.

Reading between the lines

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

The fidelity claim depends on implementation choices the reviewed text does not expose—binning width, column ordering, and the number of autoregressive passes—so a fixed, non-tuned configuration across many datasets would be the strongest test of whether the approach generalizes.
The reported membership-inference robustness is an aggregate-level result; auditing rare subpopulations or outlier records could reveal higher disclosure risk than headline attack numbers show.
Because the model is built from per-column conditional classifiers, it could plausibly be trained with a differential-privacy mechanism such as DP-SGD, offering a cheaper route to provably private tabular synthesis than diffusion-based alternatives; this is a testable extension, not a claim in the paper.

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.

Desk Editor's Note

read the letter

Colleague,

Quick take: the paper is a plausible incremental idea — discretization-based autoregressive network for mixed-type tabular data, positioned against NADE and text-as-sequence transformers, with a MIA-based privacy evaluation aimed at the DCR critique. That framing is sensible and the intro reads cleanly. But the manuscript we have contains no method and no experiments. The text ends after the first paragraph of Related Work; what follows is a corrupted reference list. So the abstract's claims about fidelity, utility, detection robustness, and privacy-utility balance are assertions, not findings.

What is actually new, as far as I can tell: extending NADE-style autoregressive density estimation to multi-categorical and mixed-type columns via discretization, and doing privacy evaluation with systematic MIAs rather than DCR. That's a legitimate program, and the related work citation set is appropriate (Uria, Shokri, van Breugel, Yao et al.). They also promise open-source code, which is welcome.

Now the soft spots, in proportion. The biggest is that the load-bearing premise — a fixed-order autoregressive factorization over discretized bins can match diffusion and GAN baselines on mixed tabular data — is neither argued nor tested in the manuscript before us. The stress-test note is correct: bin granularity and column ordering are free parameters that can dominate outcomes, and we don't see how they were chosen or whether results are robust to them. The MIA protocol is likewise absent, so the privacy conclusion rests on an unnamed attack setup. Given the authors are MOSTLY AI employees and the evaluation is self-performed, independent validation or at least a detailed protocol is essential. These are not manufactured flaws; they are simply what is missing.

I would not send this shell to peer review. There's not enough substance for a referee to engage with. The appropriate step is to ask the authors for the complete paper — method section, hyperparameters, baseline tables, error bars, and attack details — and then make a decision. If the full version delivers what the abstract promises, it could be a useful incremental contribution; nothing in the framing is incoherent.

Referee Report

4 major / 4 minor

Summary. The paper introduces TabularARGN, an auto-regressive neural network for synthetic tabular data generation, claiming high fidelity, computational efficiency, competitive performance against diffusion/GAN/transformer baselines, and strong privacy through membership-inference attack evaluation. The submitted text contains an abstract, introduction, related work, and a reference list, but no description of the TabularARGN architecture, no formal problem statement, no training objective, no experimental protocols, no result tables, no error bars, and no privacy attack details. All empirical and privacy claims are therefore assertions without supporting evidence in the reviewed manuscript.

Significance. If fully substantiated, a computationally efficient auto-regressive tabular generator with competitive fidelity and a careful MIA-based privacy evaluation would be a useful addition to the synthetic-data literature. The stated decision to move beyond DCR-type metrics in favor of systematic membership-inference attacks engages with an important current debate. The paper also promises open-source release, which would aid reproducibility. However, none of these contributions are currently checkable: the method is not specified, the experiments are not present, and the privacy analysis is not described. The significance is therefore entirely conditional.

major comments (4)

[Abstract / Section 1] The central technical claim—that 'discretization-based auto-regressive' modeling of mixed-type tabular data reaches competitive fidelity—is unsupported because no model specification appears anywhere in the manuscript. There is no definition of the factorization, the per-column discretization scheme, bin counts, column ordering, network capacity, training objective, or sampling procedure. Section 2 discusses NADE's limitations but does not state how TabularARGN overcomes them. This is not a local omission; it is the core contribution.
[Abstract / Section 1] The paper promises evaluations on statistical similarity, machine learning utility, detection robustness, and membership-inference attacks, but no experimental section, dataset list, baseline list, tables, figures, or error bars are included. Claims of 'competitive results' and an 'effective privacy-utility balance' therefore have no empirical basis in the submitted text. As a reviewer I cannot distinguish the reported performance from favorable configuration choices.
[Section 1 / Privacy evaluation] The privacy claim rests entirely on an absent MIA protocol. The paper cites [66] to argue that DCR is unreliable, but does not describe the attack model, the attacker knowledge, the training/test splits, the MIA metric, or the confidence intervals used to establish 'robustness.' Without this, the stated 'in-depth privacy evaluation' is a promise, not a result, and the reader cannot evaluate the privacy-utility trade-off.
[Section 1 (open-source claim)] The manuscript states that open-source access to TabularARGN is provided, but no repository name, URL, or artifact identifier appears. Since reproducibility is explicitly claimed as a motivation, this is a concrete, checkable omission that currently blocks verification of the method and experiments.

minor comments (4)

[References] Several reference entries are malformed or contain non-text artefacts, e.g. around [14], [20], [23], and [49]. The bibliography needs a full pass for encoding, journal/conference names, and page/article numbers.
[Throughout] Spelling is inconsistent between British ('anonymisation') and US ('anonymization') forms; please harmonize.
[Section 1] The sentence 'Synthetic data tools fundamentally act as density estimators' would benefit from a citation or formal definition of the target distribution, especially because the later privacy discussion contrasts density estimation with disclosure control.
[Section 2] The related-work discussion cites [47] on any-order autoregressive models and [65] on XLNet, but does not explain why TabularARGN adopts a fixed order. If a fixed order is a design choice, the manuscript should state it and justify it; if it is a limitation, that should be acknowledged.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reviewed text contains no equations, fitted parameters, or self-cited load-bearing results that reduce to the paper's own claims.

full rationale

The reviewed text is largely an abstract, introduction, related-work paragraph, and reference list; the method and results sections are absent. Consequently, there is no derivation chain, no equations, and no fitted parameter that could be shown by construction to equal a prediction. The central fidelity and privacy claims are empirical assertions to be checked against external baselines (e.g., diffusion, GAN, and transformer methods), so they are not defined in terms of TabularARGN's own outputs. The sole cited architectural foundation, NADE [54], is prior external work by Uria et al., not by the present authors, and it is used only as background, not as a self-citation that forces the paper's conclusions. The paper's rejection of DCR relies on an external citation [66] with no author overlap with Sidorenko and Tiwald. No uniqueness theorem from the authors' prior work is invoked, no ansatz is smuggled in via self-citation, and no known result is simply renamed. The omission of method and results is a serious completeness and reproducibility concern, but under the hard rule that circularity must be exhibited by a specific quoted reduction, the available text does not support any circularity finding. The garbled inserted passage ('Second. ... (2023)') is too degraded to identify a concrete circular step. Score 0.

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

Everything the central claim rests on: per-column binning granularity, the auto-regressive column ordering, and network/training hyperparameters are free choices the paper could tune to its benchmarks; without the method section they cannot be audited. The auto-regressive factorization, the sufficiency of discretization, and the use of MIA as the privacy yardstick are domain assumptions imported from the field. TabularARGN itself is a software architecture, not an invented entity in the sense of a new physical construct, so none are listed. No parameter-free derivation or uniqueness theorem appears in the visible text.

free parameters (3)

Per-column discretization granularity (bin count and range boundaries)
The abstract says the method 'explicitly exploits structured value ranges and discretization'; bin placement and count are per-column choices that control capacity vs. information loss. They are tuned or hand-set, and the method section containing them is not reviewable.
Auto-regressive column order
Factorizing the joint as a chain of conditionals requires a fixed column ordering; results and inductive bias change with the order. This is a modeling choice, not derived from first principles.
Network capacity and training hyperparameters
Hidden sizes, learning rates, epochs, and regularization are standard free choices in neural density estimation; they affect all reported fidelity and privacy numbers and are not visible in the reviewed text.

assumptions (4)

domain assumption The joint distribution of a tabular dataset can be factorized as a product of per-column conditional distributions without unacceptable loss.
Abstract: 'approximating the joint distribution of tabular data as a sequence of conditional probabilities.' Standard auto-regressive modeling premise; ordering and conditioning scope are choices, not theorems.
domain assumption Discretizing numeric columns into value ranges preserves enough information for statistical similarity, ML utility, and detection robustness.
Central to the fidelity claim; binning is lossy, and the abstract provides no argument that the loss is acceptable for all benchmark datasets.
domain assumption Membership-inference attack success is a valid proxy for privacy risk in synthetic data.
The privacy evaluation relies on MIAs [18, 25, 44, 48, 55, 62]; the paper itself cites [66] to reject DCR, so MIA is the chosen yardstick. Attack-setup choices can inflate or deflate the measured risk.
standard math Neural networks trained by gradient descent converge to useful density estimates on tabular data.
Implicit background for any deep generative model; not proved in the paper and not specific to TabularARGN.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Privacy-Preserving Tabular Synthetic Data Generation Using TabularARGN." pith.science (2026). https://pith.science/paper/YI3UMZNK

@misc{pith2026250806647,
  author       = {Pith},
  title        = {Pith review of: Privacy-Preserving Tabular Synthetic Data Generation Using TabularARGN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YI3UMZNK}},
  note         = {Machine review of arXiv:2508.06647}
}

read the original abstract

Synthetic data generation has become essential for securely sharing and analyzing sensitive data sets. Traditional anonymization techniques, however, often fail to adequately preserve privacy. We introduce the Tabular Auto-Regressive Generative Network (TabularARGN), a neural network architecture specifically designed for generating high-quality synthetic tabular data. Using a discretization-based auto-regressive approach, TabularARGN achieves high data fidelity while remaining computationally efficient. We evaluate TabularARGN against existing synthetic data generation methods, showing competitive results in statistical similarity, machine learning utility, and detection robustness. We further perform an in-depth privacy evaluation using systematic membership-inference attacks, highlighting the robustness and effective privacy-utility balance of our approach.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 37 canonical work pages

[66]

Zexi Yao, Nataša Krčo, Georgi Ganev, and Yves-Alexandre de Montjoye

work page
[21]

Noah Hollmann, Samuel Müller, Lennart Purucker, Katharina Eggensperger, and Frank Hutter. 2025. Accurate predictions on small data with a tabular foundation model. ��637 (2025), 319–326. https://doi.org/10.1038/s41586-024-08328-6

doi:10.1038/s41586-024-08328-6 2025
[22]

Cohen, Owen Daniel, Andrew Elliott, James Geddes, Callum Mole, Camila Rangel-Smith, and Lukasz Szpruch

Florimond Houssiau, James Jordon, Samuel N. Cohen, Owen Daniel, Andrew Elliott, James Geddes, Callum Mole, Camila Rangel-Smith, and Lukasz Szpruch

work page
[23]

Jiri Hradec, Massimo Craglia, Margherita Di Leo, Sarah De Nigris, Nicole Ost- laender, and Nicholas Nicholson. 2022. Multipurpose synthetic population for policy applications. �� (2022). https://doi.org/10.2760/50072

work page doi:10.2760/50072 2022
[24]

Yuzheng Hu, Fan Wu, Qinbin Li, Yunhui Long, Gonzalo Munilla Garrido, Chang Ge, Bolin Ding, David Forsyth, Bo Li, and Dawn Song. 2023. SoK: Privacy- Preserving Data Synthesis. arXiv:2307.02106 [cs.CR]

arXiv 2023
[25]

Jihyeon Hyeong, Jayoung Kim, Noseong Park, and Sushil Jajodia. 2022. An Em- pirical Study on the Membership Inference Attack against Tabular Data Synthesis Models. arXiv:2208.08114 [cs.CR]

work page Pith review arXiv 2022
[26]

Alexia Jolicoeur-Martineau, Kilian Fatras, and Tal Kachman. 2024. Generating and imputing tabular data via diffusion and flow-based gradient-boosted trees. In �� . PMLR, 1288–1296

work page 2024
[27]

James Jordon, Lukasz Szpruch, Florimond Houssiau, Mirko Bottarelli, Giovanni Cherubin, Carsten Maple, Samuel N Cohen, and Adrian Weller. 2022. Synthetic Data–what, why and how? �� (2022)

work page 2022

Show all 57 references

[28]

Hojjat Karami, David Atienza, and Anisoara Ionescu. 2024. SynEHRgy: Synthe- sizing Mixed-Type Structured Electronic Health Records using Decoder-Only Transformers. arXiv:2411.13428

2024 arXiv
[29]

Jayoung Kim, Chaejeong Lee, and Noseong Park. 2023. StaSy: Score-based tabular data synthesis. In ��

2023
[30]

Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv:1312.6114

2013 arXiv
[31]

Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. 2023. Tabddpm: Modelling tabular data with diffusion models. In �� . PMLR, 17564–17579

2023
[32]

Johan Leduc and Nicolas Grislain. 2021. Composable Generative Models. arXiv:2102.09249 [cs.LG]

2021 arXiv
[33]

Chaejeong Lee, Jayoung Kim, and Noseong Park. 2023. Codi: Co-evolving con- trastive diffusion models for mixed-type tabular synthesis. In �� . PMLR, 18940–18956

2023
[34]

Jin Li, Benjamin J Cairns, Jingsong Li, and Tingting Zhu. 2023. Generating syn- thetic mixed-type longitudinal electronic health records for artificial intelligent applications. �� 6, 1 (2023), 98

2023
[35]

Tennison Liu, Zhaozhi Qian, Jeroen Berrevoets, and Mihaela van der Schaar. 2023. GOGGLE: Generative modelling for tabular data by learning relational structure. In ��

2023
[36]

Junwei Ma, Apoorv Dankar, George Stein, Guangwei Yu, and Anthony Caterini. 2024. TabPFGen – Tabular Data Generation with TabPFN. arXiv:2406.05216 [cs.LG]

2024 arXiv
[37]

Calvin McCarter. 2024. Unmasking trees for tabular data. arXiv:2407.05593

2024 arXiv
[38]

Ryan McKenna, Gerome Miklau, and Daniel Sheldon. 2022. AIM: Anonymization through Microaggregation. �� 15, 11 (2022), 2422–2435

2022
[39]

Matthieu Meeus, Florent Guepin, Ana-Maria CreŢu, and Yves-Alexandre de Montjoye. 2024. Achilles’ Heels: Vulnerable Record Identification in Synthetic Data Publishing. In�� , Gene Tsudik, Mauro Conti, Kaitai Liang, and Georgios Smaragdakis (Eds.). S...

2024
[40]

Marko Miletic and Murat Sariyar. 2024. Assessing the Potentials of LLMs and GANs as State-of-the-Art Tabular Synthetic Data Generation Methods. In�� , Josep Domingo-Ferrer and Melek Önen (Eds.). Springer Nature Switzerland, Cham, 374–389

2024
[41]

Markus Mueller. 2025. Continuous Diffusion for Mixed-Type Tabular Data. https://github.com/muellermarkus/cdtd

2025
[42]

Markus Mueller, Kathrin Gruber, and Dennis Fok. 2025. Continuous Diffusion for Mixed-Type Tabular Data. arXiv:2312.10431 [cs.LG]

2025
[43]

Noseong Park, Mahmoud Mohammadi, Kshitij Gorde, Sushil Jajodia, Hongkyu Park, and Youngmin Kim. 2018. Data synthesis based on generative adversarial networks. arXiv:1806.03384

2018 arXiv
[44]

Lisa Pilgram, Fida K. Dankar, Jorg Drechsler, Mark Elliot, Josep Domingo-Ferrer, Paul Francis, Murat Kantarcioglu, Linglong Kong, Bradley Malin, Krishnamurty Muralidhar, Puja Myles, Fabian Prasser, Jean Louis Raisaro, Chao Yan, and Khaled El Emam. 2025. A Consensus Privacy Met...

2025 arXiv
[45]

Michael Platzer and Thomas Reutterer. 2021. Holdout-based empirical assessment of mixed-type synthetic data. �� 4 (2021), 679939

2021
[46]

Zhaozhi Qian, Bogdan-Constantin Cebere, and Mihaela van der Schaar. 2023. Synthcity: facilitating innovative use cases of synthetic data in different data modalities, 2023. https://doi.org/10.48550/arXiv

2023 doi
[47]

Andy Shih, Dorsa Sadigh, and Stefano Ermon. 2022. Training and Inference on Any-Order Autoregressive Models the Right Way. arXiv:2205.13554 [cs.LG]

2022 arXiv
[48]

Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. 2017. Mem- bership Inference Attacks Against Machine Learning Models . In �� . IEEE Computer Society, Los Alamitos, CA, USA, 3–18. https://doi.org/10.1109/SP.2017.41

2017 doi
[49]

Daniel Smolyak, Margrét V Bjarnadóttir, Kenyon Crowley, and Ritu Agarwal
[50]

Aivin V Solatorio and Olivier Dupriez. 2023. Realtabformer: Generating realistic relational and tabular data using transformers. arXiv:2302.02041

2023 arXiv
[51]

Theresa Stadler, Bristena Oprisanu, and Carmela Troncoso. 2022. Synthetic Data – Anonymisation Groundhog Day. In �� . USENIX Association, Boston, MA, 1451–1468

2022
[52]

Amy Steier, Lipika Ramaswamy, Andre Manoel, and Alexa Haushalter. 2025. Synthetic Data Privacy Metrics. arXiv:2501.03941 [cs.LG]

2025 arXiv
[53]

UNECE. 2023. �� . United Nations. https: //www.un-ilibrary.org/content/books/9789210021708

2023
[54]

Benigno Uria, Marc-Alexandre Côté, Karol Gregor, Iain Murray, and Hugo Larochelle. 2016. Neural autoregressive distribution estimation. �� 17, 205 (2016), 1–37

2016
[55]

Boris van Breugel, Hao Sun, Zhaozhi Qian, and Mihaela van der Schaar. 2023. Membership Inference Attacks against Synthetic Data through Overfitting De- tection. arXiv:2302.12580 [cs.LG]

2023 arXiv
[56]

Boris van Breugel and Mihaela van der Schaar. 2023. Beyond Privacy: Navigating the Opportunities and Challenges of Synthetic Data. arXiv:2304.03722 [cs.LG]

2023 arXiv
[57]

Mihaela van der Schaar, Boris van Breugel, Trent Kyono, and Jeroen Berrevoets
[58]

Giuseppe Vietri, Cedric Archambeau, Sergul Aydore, William Brown, Michael Kearns, Aaron Roth, Ankit Siva, Shuai Tang, and Steven Z. Wu. 2022. Pri- vate Synthetic Data for Multitask Learning and Marginal Queries. In �� , S. Koyejo...

2022
[59]

Mario Villaizán-Vallelado, Matteo Salvatori, Carlos Segura, and Ioannis Ara- pakis. 2024. Diffusion Models for Tabular Data Imputation and Synthetic Data Generation. arXiv:2407.02549 [cs.LG]

2024 arXiv
[60]

Yijue Wang, Chenghong Wang, Zigeng Wang, Shanglin Zhou, Hang Liu, Jinbo Bi, Caiwen Ding, and Sanguthevar Rajasekaran. 2021. Against Membership Inference Attack: Pruning is All You Need. In�� ...

2021 doi
[61]

Watson, Kristin Blesch, Jan Kapar, and Marvin N

David S. Watson, Kristin Blesch, Jan Kapar, and Marvin N. Wright. 2023. Adversarial random forests for density estimation and generative modeling. arXiv:2205.09435 [stat.ML]

2023 arXiv
[62]

Feng Wu, Lei Cui, Shaowen Yao, and Shui Yu. 2024. Inference Attacks: A Taxon- omy, Survey, and Promising Directions. arXiv:2406.02027 [cs.LG]

2024 arXiv
[63]

Liyang Xie, Kaixiang Lin, Shu Wang, Fei Wang, and Jiayu Zhou. 2018. Differen- tially Private Generative Adversarial Network. arXiv:1802.06739 [cs.LG]

2018 arXiv
[64]

Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni
[65]

Zhilin Yang. 2019. XLNet: Generalized Autoregressive Pretraining for Language Understanding. arXiv:1906.08237

2019 arXiv
[67]

Jinsung Yoon, James Jordon, and Mihaela van der Schaar. 2019. PATE-GAN: Generating Synthetic Data with Differential Privacy Guarantees. In ��

2019
[68]

Hengrui Zhang, Liancheng Fang, Qitian Wu, and Philip S. Yu. 2024. Diffusion-nested Auto-Regressive Synthesis of Heterogeneous Tabular Data. arXiv:2410.21523 [cs.LG]

2024 arXiv
[69]

Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan, Zhengyuan Shen, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. 2024. Mixed- Type Tabular Data Synthesis with Score-based Diffusion in Latent Space. In �� ...

2024
[70]

Zilong Zhao, Aditya Kunar, Robert Birke, and Lydia Y Chen. 2021. Ctab-gan: Effective table data synthesizing. In�� . PMLR, 97–112

2021
[71]

Zilong Zhao, Aditya Kunar, Robert Birke, Hiek Van der Scheer, and Lydia Y Chen
[72]

�� 6 (2024), 1296508

Ctab-gan+: Enhancing tabular data synthesis. �� 6 (2024), 1296508. 12

2024
[2019]

Modeling tabular data using conditional gan.�� 32 (2019)

2019
[2021]

In �� , M

DECAF: Generating Fair Synthetic Data Using Causally-Aware Generative Networks. In �� , M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan (Eds.), Vol. 34. Curran Associates, Inc., 22221–22233
[2022]

arXiv:2211.06550 [cs.CR]

TAPAS: a Toolbox for Adversarial Privacy Auditing of Synthetic Data. arXiv:2211.06550 [cs.CR]

arXiv
[2024]

�� 7, 4 (2024), ooae114

Large language models and synthetic health data: progress and prospects. �� 7, 4 (2024), ooae114

2024
[2025]

arXiv:2505.01524 [cs.CR]

The DCR Delusion: Measuring the Privacy Risk of Synthetic Data. arXiv:2505.01524 [cs.CR]

arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [66]

Zexi Yao, Nataša Krčo, Georgi Ganev, and Yves-Alexandre de Montjoye

work page

[2] [21]

Noah Hollmann, Samuel Müller, Lennart Purucker, Katharina Eggensperger, and Frank Hutter. 2025. Accurate predictions on small data with a tabular foundation model. ��637 (2025), 319–326. https://doi.org/10.1038/s41586-024-08328-6

doi:10.1038/s41586-024-08328-6 2025

[3] [22]

Cohen, Owen Daniel, Andrew Elliott, James Geddes, Callum Mole, Camila Rangel-Smith, and Lukasz Szpruch

Florimond Houssiau, James Jordon, Samuel N. Cohen, Owen Daniel, Andrew Elliott, James Geddes, Callum Mole, Camila Rangel-Smith, and Lukasz Szpruch

work page

[4] [23]

Jiri Hradec, Massimo Craglia, Margherita Di Leo, Sarah De Nigris, Nicole Ost- laender, and Nicholas Nicholson. 2022. Multipurpose synthetic population for policy applications. �� (2022). https://doi.org/10.2760/50072

work page doi:10.2760/50072 2022

[5] [24]

Yuzheng Hu, Fan Wu, Qinbin Li, Yunhui Long, Gonzalo Munilla Garrido, Chang Ge, Bolin Ding, David Forsyth, Bo Li, and Dawn Song. 2023. SoK: Privacy- Preserving Data Synthesis. arXiv:2307.02106 [cs.CR]

arXiv 2023

[6] [25]

Jihyeon Hyeong, Jayoung Kim, Noseong Park, and Sushil Jajodia. 2022. An Em- pirical Study on the Membership Inference Attack against Tabular Data Synthesis Models. arXiv:2208.08114 [cs.CR]

work page Pith review arXiv 2022

[7] [26]

Alexia Jolicoeur-Martineau, Kilian Fatras, and Tal Kachman. 2024. Generating and imputing tabular data via diffusion and flow-based gradient-boosted trees. In �� . PMLR, 1288–1296

work page 2024

[8] [27]

James Jordon, Lukasz Szpruch, Florimond Houssiau, Mirko Bottarelli, Giovanni Cherubin, Carsten Maple, Samuel N Cohen, and Adrian Weller. 2022. Synthetic Data–what, why and how? �� (2022)

work page 2022