PATE-TabTransGAN: Differentially Private Synthetic Tabular Data Generation via Transformer-Based Student Discrimination

M. Wo\'zniak; M. Youssef

arxiv: 2605.26802 · v1 · pith:OKA3233Inew · submitted 2026-05-26 · 💻 cs.LG

PATE-TabTransGAN: Differentially Private Synthetic Tabular Data Generation via Transformer-Based Student Discrimination

M. Youssef , M. Wo\'zniak This is my paper

Pith reviewed 2026-06-29 19:30 UTC · model grok-4.3

classification 💻 cs.LG

keywords differentially private synthetic datatabular data generationPATE mechanismTransformer discriminatorGANresidual generatorprivacy accounting

0 comments

The pith

PATE-TabTransGAN pairs a PATE teacher ensemble with a Transformer student discriminator to generate formally private synthetic tabular data that matches or exceeds baselines on AUROC.

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

The paper presents a generative framework that trains an ensemble of logistic regression teachers on data partitions and supplies noisy aggregated labels to a Transformer-based student discriminator inside a GAN. The residual generator is then optimized against this student, inheriting (ε, δ)-DP guarantees through post-processing while the Transformer architecture models column dependencies. On the Adult, Breast, Cardio, and Cervical benchmarks the method records the best or tied-best AUROC against PATE-GAN, DP-GAN, and DP-CTGAN, with AUCPR results that are competitive once evaluation conventions are accounted for.

Core claim

PATE-TabTransGAN integrates the Private Aggregation of Teacher Ensembles mechanism with a Transformer-based student discriminator and GNMax RDP accounting; the resulting student supplies a differentially private training signal to a residual generator, producing synthetic tabular data that attains the best or tied-best AUROC on all four tested datasets while satisfying formal privacy.

What carries the argument

The Transformer student discriminator trained on noisy PATE-aggregated labels, which transfers formal differential privacy to the generator by post-processing.

If this is right

Downstream models trained on the synthetic tables inherit formal privacy protection without additional noise injection.
The residual generator can be swapped for other architectures while the privacy guarantee remains intact by post-processing.
AUCPR sensitivity to class-label convention implies that utility comparisons across pipelines require explicit alignment of evaluation rules.
The GNMax accountant enables numerically stable tracking of privacy loss across multiple teacher queries.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Replacing logistic regression teachers with more expressive models could increase label signal and further improve downstream AUROC.
The same PATE-student pattern could be tested on sequential or graph-structured tabular data to check whether the Transformer advantage generalizes.
Tighter privacy accounting or adaptive teacher partitioning might reduce the noise level required for a target ε without changing the student architecture.

Load-bearing premise

The noisy labels supplied by the PATE teacher ensemble still contain enough signal for the Transformer student to learn inter-feature dependencies.

What would settle it

On the same four datasets, a re-evaluation that uses an identical positive-class convention for Adult would remove the reported AUCPR gap if that gap is caused only by convention rather than by synthesis quality.

Figures

Figures reproduced from arXiv: 2605.26802 by M. Wo\'zniak, M. Youssef.

**Figure 1.** Figure 1: Overview of the PATE-TabTransGAN architecture. An ensemble of k Logistic Regression teachers, each trained on a disjoint data shard, votes on generator-produced samples; votes are aggregated with Gaussian noise to produce private labels that supervise the Transformer-based student discriminator and the residual generator. The trained generator inherits a formal (ε, δ)-DP guarantee by the post-processing p… view at source ↗

**Figure 2.** Figure 2: Effect of positive-class convention on Adult AUROC and AUCPR at ε = 3. Before: minority class (income >$50K, ≈ 24%) as positive (our default). After : majority class as positive, matching the convention implied by [8]. Reproduced DP-CTGAN values shown in both panels. Scope and limitations. The re-evaluation shows the conventional difference is sufficient to explain the observed AUCPR gap; this does not es… view at source ↗

**Figure 3.** Figure 3: Per-classifier AUROC (left) and AUCPR (right) for PATE-TabTransGAN across the four datasets, mean over five runs. Outlined cells mark the best classifier per row. Utility under matched privacy. While the privacy-utility trade-off remains a fundamental challenge for private data synthesis, PATE-TabTransGAN significantly narrows this gap. Under matched privacy budgets, it attains the highest mean AUROC on t… view at source ↗

read the original abstract

Generating high-fidelity synthetic tabular data under formal differential privacy guarantees remains an open challenge. Methods that provide strong theoretical protection typically sacrifice the modeling of inter-feature dependencies required for realistic synthesis, while architectures that excel at capturing complex column relationships offer only empirical privacy guarantees. We present PATE-TabTransGAN, a generative framework that integrates the Private Aggregation of Teacher Ensembles (PATE) mechanism with a Transformer-based student discriminator to jointly address both requirements, and employs a GNMax RDP accountant for numerically stable privacy accounting. An ensemble of Logistic Regression teachers trained on disjoint partitions supervise the student via noisy-aggregated labels, and a residual generator is optimized against this differentially private student, inheriting formal ({\epsilon}, {\delta})-DP guarantees by post-processing. PATE-TabTransGAN was compared with PATE-GAN, DP-GAN, and DP-CTGAN, considered state-of-the-art in differentially private tabular synthesis. Experiments conducted on four tabular benchmarks (Adult, Breast, Cardio, Cervical) confirmed the high quality of the proposed method: PATE-TabTransGAN attains the best or tied-best AUROC on all four datasets. On AUCPR it matches the strongest baseline on Cardio, leads on Cervical, and trails on Breast; on Adult, we demonstrate that AUCPR is highly sensitive to positive-class convention, and that the observed gap is consistent with a convention difference between evaluation pipelines rather than a synthesis deficit.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PATE-TabTransGAN is a new PATE-plus-Transformer-discriminator combo inside a residual GAN that reports best or tied AUROC on four tabular benchmarks, but the abstract leaves variance, hyper-parameters, and privacy accounting thin.

read the letter

The main takeaway is that this paper puts a Transformer discriminator into the PATE student slot of a residual GAN for differentially private tabular synthesis and gets the best or tied AUROC on Adult, Breast, Cardio, and Cervical against PATE-GAN, DP-GAN, and DP-CTGAN. The architecture itself is a fresh pairing not shown in the cited baselines.

What works is the clean integration: logistic regression teachers on disjoint data, GNMax noisy aggregation, and post-processing to inherit (ε, δ)-DP. The experiments are run on standard benchmarks and the AUROC wins are stated directly. The note on AUCPR sensitivity to positive-class convention is also useful and shows the authors paid attention to evaluation details.

The soft spots are exactly where the abstract is thin. No error bars appear, no full hyper-parameter table is referenced, and the privacy accounting is described at a high level without the numerical RDP values or sensitivity analysis. The stress-test concern lands: the teachers are linear models, so their votes carry only linear boundaries plus GNMax noise. Nothing in the construction proves the Transformer can recover higher-order feature dependencies from that signal, yet the claimed advantage rests on the Transformer's capacity for exactly those dependencies. The results show the method is competitive, but without ablation on teacher type or noise level it is hard to tell how much the architecture actually drives the gains versus implementation choices.

This paper is for groups already working on DP tabular synthesis who want to try the PATE-Transformer student idea. It is not a broad theoretical advance, but the concrete architecture and the reported numbers are solid enough to merit referee time. I would send it to review.

Referee Report

3 major / 2 minor

Summary. The paper introduces PATE-TabTransGAN, a framework integrating PATE (using an ensemble of Logistic Regression teachers on disjoint partitions) with a Transformer-based student discriminator supervised via GNMax-noisy aggregated labels; a residual generator is then trained against this student to produce synthetic tabular data inheriting (ε, δ)-DP guarantees by post-processing. It reports that the method attains the best or tied-best AUROC on all four benchmarks (Adult, Breast, Cardio, Cervical) versus PATE-GAN, DP-GAN, and DP-CTGAN, with mixed AUCPR results and a note on AUCPR sensitivity to positive-class convention.

Significance. If the empirical results hold under rigorous validation, the work would demonstrate a viable route to formal DP tabular synthesis that leverages Transformer capacity for inter-feature dependencies while using GNMax for stable RDP accounting; this addresses a key tension between privacy theory and modeling power. The explicit use of GNMax for numerically stable privacy accounting is a concrete technical strength that aids reproducibility of the guarantees.

major comments (3)

[Abstract; Experiments section] Abstract and experimental results: the central claim of best/tied-best AUROC on all four datasets is reported without error bars, standard deviations across runs, or statistical significance tests. This directly weakens confidence in whether the observed wins are robust or attributable to the claimed architecture.
[Method description (PATE integration and student discriminator)] Method (PATE teacher-student setup): the student is a Transformer trained on noisy labels from linear LR teachers, yet no ablation, dependency analysis, or diagnostic is provided showing that higher-order column relationships survive the linear supervision plus GNMax perturbation. This assumption is load-bearing for attributing gains to the Transformer rather than implementation details.
[Abstract; Privacy accounting subsection] Privacy accounting: although GNMax RDP is invoked for formal guarantees, the manuscript supplies neither the concrete noise multiplier, teacher count, resulting (ε, δ) values, nor the full accounting trace for the reported experiments. These details are required to substantiate the post-processing DP claim.

minor comments (2)

[Experiments; AUCPR analysis] The discussion of AUCPR convention sensitivity on Adult is helpful; extend it by stating the exact positive-class convention applied to every baseline for transparency.
[Experimental setup] Hyper-parameter choices (e.g., number of teachers, noise scale, Transformer depth) should be collected in a single table rather than scattered in text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point-by-point below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: [Abstract; Experiments section] Abstract and experimental results: the central claim of best/tied-best AUROC on all four datasets is reported without error bars, standard deviations across runs, or statistical significance tests. This directly weakens confidence in whether the observed wins are robust or attributable to the claimed architecture.

Authors: We agree that the absence of error bars and statistical tests reduces confidence in the robustness of the reported AUROC improvements. In the revised manuscript we will report means and standard deviations over multiple independent runs (with fixed seeds) and include paired statistical significance tests (e.g., Wilcoxon or t-tests) against the baselines to substantiate the claims. revision: yes
Referee: [Method description (PATE integration and student discriminator)] Method (PATE teacher-student setup): the student is a Transformer trained on noisy labels from linear LR teachers, yet no ablation, dependency analysis, or diagnostic is provided showing that higher-order column relationships survive the linear supervision plus GNMax perturbation. This assumption is load-bearing for attributing gains to the Transformer rather than implementation details.

Authors: The linear teachers supply only noisy binary labels; the Transformer student still receives the full feature vectors and must learn a decision boundary that captures higher-order interactions to minimize the discrimination loss. Nevertheless, we acknowledge that an explicit ablation would strengthen attribution. We will add a controlled comparison of Transformer versus linear student discriminators (keeping teachers and GNMax fixed) to quantify the contribution of non-linear capacity. revision: yes
Referee: [Abstract; Privacy accounting subsection] Privacy accounting: although GNMax RDP is invoked for formal guarantees, the manuscript supplies neither the concrete noise multiplier, teacher count, resulting (ε, δ) values, nor the full accounting trace for the reported experiments. These details are required to substantiate the post-processing DP claim.

Authors: We will include the exact experimental parameters (number of teachers, noise multiplier σ, and the resulting (ε, δ) values) together with the complete GNMax RDP accounting trace for each dataset in a new subsection of the revised manuscript, ensuring full reproducibility of the privacy guarantees. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation or performance claims

full rationale

The paper describes an empirical construction (PATE ensemble of LR teachers + Transformer student + residual generator + GNMax accountant) and reports AUROC/AUCPR from direct experimental comparison against external baselines on four public datasets. No equations, derivations, or fitted quantities are shown that reduce the reported metrics to internal definitions or self-citations by construction. The central claims rest on external benchmark results rather than any self-referential reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The work rests on standard differential privacy theory and the existing PATE construction; no new mathematical entities are postulated and the only free parameters are conventional hyper-parameters such as teacher count and noise scale.

free parameters (2)

number of teachers
Controls the ensemble size in the PATE mechanism and is chosen per experiment.
noise multiplier for GNMax
Determines the privacy budget spent during label aggregation.

axioms (2)

standard math Differential privacy post-processing property
Invoked to transfer (ε,δ)-DP from the student to the generator.
standard math Rényi DP composition via GNMax accountant
Used for numerically stable tracking of cumulative privacy loss.

pith-pipeline@v0.9.1-grok · 5794 in / 1257 out tokens · 51500 ms · 2026-06-29T19:30:01.397821+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 4 canonical work pages · 4 internal anchors

[1]

In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security

Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security. pp. 308– 318 (2016)

2016
[2]

IEEE Transactions on Knowledge and Data Engineering19(11), 1450–1464 (2007)

Angiulli, F.: Fast nearest neighbor condensation for large data sets classification. IEEE Transactions on Knowledge and Data Engineering19(11), 1450–1464 (2007)

2007
[3]

In: The Eleventh International Conference on Learning Representations (2022) PATE-TabTransGAN 15

Carlini, N., Ippolito, D., Jagielski, M., Lee, K., Tramer, F., Zhang, C.: Quantify- ing memorization across neural language models. In: The Eleventh International Conference on Learning Representations (2022) PATE-TabTransGAN 15

2022
[4]

In: 30th USENIX security symposium (USENIX Se- curity 21)

Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al.: Extracting training data from large language models. In: 30th USENIX security symposium (USENIX Se- curity 21). pp. 2633–2650 (2021)

2021
[5]

Communications of the ACM54(1), 86–95 (2011)

Dwork, C.: A firm foundation for private data analysis. Communications of the ACM54(1), 86–95 (2011)

2011
[6]

In: Theory of Cryptography Conference

Dwork, C., McSherry, F., Nissim, K., Smith, A.: Calibrating noise to sensitivity in private data analysis. In: Theory of Cryptography Conference. pp. 265–284. Springer (2006)

2006
[7]

Founda- tions and trends®in theoretical computer science9(3-4), 211–487 (2014)

Dwork, C., Roth, A.: The algorithmic foundations of differential privacy. Founda- tions and trends®in theoretical computer science9(3-4), 211–487 (2014)

2014
[8]

In: International conference on artificial intelligence in medicine

Fang, M.L., Dhami, D.S., Kersting, K.: Dp-ctgan: Differentially private medical data generation using ctgans. In: International conference on artificial intelligence in medicine. pp. 178–188. Springer (2022)

2022
[9]

In: Proceedingsofthe52ndannualACMSIGACTsymposiumontheoryofcomputing

Feldman, V.: Does learning require memorization? a short tale about a long tail. In: Proceedingsofthe52ndannualACMSIGACTsymposiumontheoryofcomputing. pp. 954–959 (2020)

2020
[10]

In: Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security

Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit con- fidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security. pp. 1322–1333 (2015)

2015
[11]

In: International Conference on Machine Learning

Gaboardi, M., Arias, E.J.G., Hsu, J., Roth, A., Wu, Z.S.: Dual query: Practical private query release for high dimensional data. In: International Conference on Machine Learning. pp. 1170–1178. PMLR (2014)

2014
[12]

Advances in Neural Information Processing Systems36, 46245–46254 (2023)

Gulati, M., Roysdon, P.: Tabmt: Generating tabular data with masked transform- ers. Advances in Neural Information Processing Systems36, 46245–46254 (2023)

2023
[13]

Categorical Reparameterization with Gumbel-Softmax

Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[14]

In: International Conference on Learning Represen- tations (2018)

Jordon, J., Yoon, J., Van Der Schaar, M.: Pate-gan: Generating synthetic data with differential privacy guarantees. In: International Conference on Learning Represen- tations (2018)

2018
[15]

In: 2017 IEEE 30th computer security foun- dations symposium (CSF)

Mironov, I.: Rényi differential privacy. In: 2017 IEEE 30th computer security foun- dations symposium (CSF). pp. 263–275. IEEE (2017)

2017
[16]

Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data

Papernot, N., Abadi, M., Erlingsson, U., Goodfellow, I., Talwar, K.: Semi- supervised knowledge transfer for deep learning from private training data. arXiv preprint arXiv:1610.05755 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[17]

Scalable Private Learning with PATE

Papernot, N., Song, S., Mironov, I., Raghunathan, A., Talwar, K., Erlingsson, Ú.: Scalable private learning with pate. arXiv preprint arXiv:1802.08908 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[18]

Differentially Private Generative Adversarial Network

Xie, L., Lin, K., Wang, S., Wang, F., Zhou, J.: Differentially private generative adversarial network. arXiv preprint arXiv:1802.06739 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[19]

Advances in neural information processing systems32 (2019)

Xu,L.,Skoularidou,M.,Cuesta-Infante,A.,Veeramachaneni,K.:Modelingtabular data using conditional gan. Advances in neural information processing systems32 (2019)

2019
[20]

Information Processing & Management62(5), 104220 (2025)

Zhang, H., Jing, Y., Zhang, F., Li, Z., Wang, X.S., Chen, Z., Lv, C.: Tabtransgan: A hybrid approach integrating gan and transformer architectures for tabular data synthesis. Information Processing & Management62(5), 104220 (2025)

2025
[21]

In: 30th USENIX Security Sympo- sium (USENIX Security 21)

Zhang, Z., Wang, T., Li, N., Honorio, J., Backes, M., He, S., Chen, J., Zhang, Y.: Privsyn: Differentially private data synthesis. In: 30th USENIX Security Sympo- sium (USENIX Security 21). pp. 929–946 (2021) 16 M. Youssef and M. Woźniak

2021
[22]

In: Proceed- ings of the IEEE/CVF winter conference on applications of computer vision

Zhao, B., Bilen, H.: Dataset condensation with distribution matching. In: Proceed- ings of the IEEE/CVF winter conference on applications of computer vision. pp. 6514–6523 (2023)

2023

[1] [1]

In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security

Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security. pp. 308– 318 (2016)

2016

[2] [2]

IEEE Transactions on Knowledge and Data Engineering19(11), 1450–1464 (2007)

Angiulli, F.: Fast nearest neighbor condensation for large data sets classification. IEEE Transactions on Knowledge and Data Engineering19(11), 1450–1464 (2007)

2007

[3] [3]

In: The Eleventh International Conference on Learning Representations (2022) PATE-TabTransGAN 15

Carlini, N., Ippolito, D., Jagielski, M., Lee, K., Tramer, F., Zhang, C.: Quantify- ing memorization across neural language models. In: The Eleventh International Conference on Learning Representations (2022) PATE-TabTransGAN 15

2022

[4] [4]

In: 30th USENIX security symposium (USENIX Se- curity 21)

Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al.: Extracting training data from large language models. In: 30th USENIX security symposium (USENIX Se- curity 21). pp. 2633–2650 (2021)

2021

[5] [5]

Communications of the ACM54(1), 86–95 (2011)

Dwork, C.: A firm foundation for private data analysis. Communications of the ACM54(1), 86–95 (2011)

2011

[6] [6]

In: Theory of Cryptography Conference

Dwork, C., McSherry, F., Nissim, K., Smith, A.: Calibrating noise to sensitivity in private data analysis. In: Theory of Cryptography Conference. pp. 265–284. Springer (2006)

2006

[7] [7]

Founda- tions and trends®in theoretical computer science9(3-4), 211–487 (2014)

Dwork, C., Roth, A.: The algorithmic foundations of differential privacy. Founda- tions and trends®in theoretical computer science9(3-4), 211–487 (2014)

2014

[8] [8]

In: International conference on artificial intelligence in medicine

Fang, M.L., Dhami, D.S., Kersting, K.: Dp-ctgan: Differentially private medical data generation using ctgans. In: International conference on artificial intelligence in medicine. pp. 178–188. Springer (2022)

2022

[9] [9]

In: Proceedingsofthe52ndannualACMSIGACTsymposiumontheoryofcomputing

Feldman, V.: Does learning require memorization? a short tale about a long tail. In: Proceedingsofthe52ndannualACMSIGACTsymposiumontheoryofcomputing. pp. 954–959 (2020)

2020

[10] [10]

In: Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security

Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit con- fidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security. pp. 1322–1333 (2015)

2015

[11] [11]

In: International Conference on Machine Learning

Gaboardi, M., Arias, E.J.G., Hsu, J., Roth, A., Wu, Z.S.: Dual query: Practical private query release for high dimensional data. In: International Conference on Machine Learning. pp. 1170–1178. PMLR (2014)

2014

[12] [12]

Advances in Neural Information Processing Systems36, 46245–46254 (2023)

Gulati, M., Roysdon, P.: Tabmt: Generating tabular data with masked transform- ers. Advances in Neural Information Processing Systems36, 46245–46254 (2023)

2023

[13] [13]

Categorical Reparameterization with Gumbel-Softmax

Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[14] [14]

In: International Conference on Learning Represen- tations (2018)

Jordon, J., Yoon, J., Van Der Schaar, M.: Pate-gan: Generating synthetic data with differential privacy guarantees. In: International Conference on Learning Represen- tations (2018)

2018

[15] [15]

In: 2017 IEEE 30th computer security foun- dations symposium (CSF)

Mironov, I.: Rényi differential privacy. In: 2017 IEEE 30th computer security foun- dations symposium (CSF). pp. 263–275. IEEE (2017)

2017

[16] [16]

Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data

Papernot, N., Abadi, M., Erlingsson, U., Goodfellow, I., Talwar, K.: Semi- supervised knowledge transfer for deep learning from private training data. arXiv preprint arXiv:1610.05755 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[17] [17]

Scalable Private Learning with PATE

Papernot, N., Song, S., Mironov, I., Raghunathan, A., Talwar, K., Erlingsson, Ú.: Scalable private learning with pate. arXiv preprint arXiv:1802.08908 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[18] [18]

Differentially Private Generative Adversarial Network

Xie, L., Lin, K., Wang, S., Wang, F., Zhou, J.: Differentially private generative adversarial network. arXiv preprint arXiv:1802.06739 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[19] [19]

Advances in neural information processing systems32 (2019)

Xu,L.,Skoularidou,M.,Cuesta-Infante,A.,Veeramachaneni,K.:Modelingtabular data using conditional gan. Advances in neural information processing systems32 (2019)

2019

[20] [20]

Information Processing & Management62(5), 104220 (2025)

Zhang, H., Jing, Y., Zhang, F., Li, Z., Wang, X.S., Chen, Z., Lv, C.: Tabtransgan: A hybrid approach integrating gan and transformer architectures for tabular data synthesis. Information Processing & Management62(5), 104220 (2025)

2025

[21] [21]

In: 30th USENIX Security Sympo- sium (USENIX Security 21)

Zhang, Z., Wang, T., Li, N., Honorio, J., Backes, M., He, S., Chen, J., Zhang, Y.: Privsyn: Differentially private data synthesis. In: 30th USENIX Security Sympo- sium (USENIX Security 21). pp. 929–946 (2021) 16 M. Youssef and M. Woźniak

2021

[22] [22]

In: Proceed- ings of the IEEE/CVF winter conference on applications of computer vision

Zhao, B., Bilen, H.: Dataset condensation with distribution matching. In: Proceed- ings of the IEEE/CVF winter conference on applications of computer vision. pp. 6514–6523 (2023)

2023