Pith. sign in

REVIEW 3 major objections 5 minor 52 references

dpmm: Differentially Private Marginal Models, a Library for Synthetic Tabular Data Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read dpmm is the only DP synthetic-data library whose models ship with end-to-end privacy and no known leaks.

desk verdict A solid engineering library paper for DP synthetic data: the code and audits are genuinely useful, but the headline 'only library without known vulnerabilities' is conditional on an unenforced user assumption about the data domain, and utility claims rest on a single dataset without error bars. read the letter →

arxiv 2506.00322 v1 pith:B2LEFZS5 submitted 2025-05-31 cs.CR cs.AIcs.LG

Sofiane Mahiou , Amir Dizche , Reza Nazari , Xinmin Wu , Ralph Abbey , Jorge Silva , Georgi Ganev This is my paper
classification cs.CRcs.AIcs.LG
keywords differentialprivacysyntheticdatamarginalmodelsPrivBayesMSTAIMPrivate-PGMDPauditing
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

dpmm is an open-source library for generating synthetic tabular data under differential privacy (DP). The paper claims that dpmm is the only such library that combines the three main marginal models—PrivBayes+PGM, MST, and AIM—with an end-to-end DP pipeline that has no known privacy-related vulnerabilities. On the Wine benchmark at $\epsilon = 1$, dpmm reports average utility 1.5% higher than private-pgm and 147% higher than OpenDP/synthcity, with utility increasing as the privacy budget grows. The paper's DP auditing experiments with AuditSynth and $\Delta$-Siege show dpmm remaining within expected privacy-loss bounds while flagging general or floating-point violations in other libraries. A practitioner gets a single pip-installable tool with documented privacy handling and rich functionality such as conditional generation.

What carries the argument

The load-bearing machinery is the select-measure-generate pipeline plus three privacy safeguards. Select-measure-generate chooses the marginals of interest—PrivBayes builds a low-degree Bayesian network with the Exponential mechanism, MST picks a maximum spanning tree over pairwise correlations, and AIM chooses higher-order marginals iteratively—then measures them with the Gaussian mechanism and reconstructs synthetic records with Private-PGM. The safeguards are a user-supplied or DP-extracted data domain, DP discretization using PrivTree or uniform binning, and a floating-point-safe Gaussian sampler that avoids the precision vulnerabilities found in some numerical libraries. Private-PGM also supplies conditional generation, structural-zero preservation, and model-size control, making it the functional core of the library.

What would settle it

Run the AuditSynth membership game against dpmm models trained with a data domain derived from the private training data (for example, exact column min/max) without charging the preprocessing budget; an empirical epsilon above the theoretical epsilon would refute the claimed end-to-end guarantee. A second check is to trace the domain-handling code path and show that such a training call is accepted with zero privacy cost.

Watch

Extended reading notes

Core claim

The central claim is that, to the best of the authors' knowledge, dpmm is the only library that combines the three main marginal models—PrivBayes+PGM, MST, and AIM—with an end-to-end differential privacy guarantee and no known privacy-related vulnerabilities. The models share the select-measure-generate paradigm: each chooses a collection of marginals (via a Bayesian network, a maximum spanning tree, or adaptive iteration), measures them with the Gaussian mechanism, and reconstructs synthetic data consistent with those noisy marginals using Private-PGM. End-to-end DP rests on three mechanisms: a data domain that is public or DP-extracted, DP discretization via PrivTree or uniform binning, and floating-point-safe Gaussian sampling. The empirical evidence is a utility comparison on Wine and white-box audits with AuditSynth and $\Delta$-Siege, which keep dpmm's empirical privacy loss within expected bounds while exposing violations in some other libraries.

Load-bearing premise

The load-bearing premise is that the data domain is either public knowledge or estimated with the allocated half of the preprocessing budget, because a domain computed from sensitive data without differential privacy silently voids the end-to-end guarantee, and dpmm does not enforce this condition programmatically.

Editorial extensions

If this is right

  • dpmm gives practitioners a single pip-installable library in which PrivBayes+PGM, MST, and AIM all carry end-to-end DP guarantees.
  • Models can be trained on mixed numerical and categorical data, generate conditionally on arbitrary constraints without retraining, pretrain on public data at zero privacy cost, and preserve structural zeros.
  • The reported utility gains suggest that the gap between dpmm and some older implementations comes largely from preprocessing choices rather than from the marginal-selection algorithms themselves.
  • The audit results imply that libraries whose data domain leaks into the model cannot claim the same end-to-end guarantee, while dpmm's design closes that leak.
  • With all three models under one API, dpmm allows direct comparison of PrivBayes+PGM, MST, and AIM on the same data and privacy budget without switching implementations.

Reading between the lines

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

  • If the end-to-end claim holds beyond the Wine dataset, dpmm could become a practical drop-in component for organizations that avoid DP synthetic data because existing libraries leak through domain extraction or floating-point noise.
  • The 147% utility advantage over OpenDP/synthcity is measured on a single dataset with default hyperparameters; on datasets with fully public domains and uniform preprocessing, the advantage may narrow, since much of the gap appears tied to domain handling.
  • A natural extension would be to enforce the domain assumption programmatically: if dpmm detected that a supplied domain was derived from sensitive values, it could refuse to train or charge privacy budget, making the guarantee robust to user error.
  • The same auditing method could test other libraries: patching a floating-point-safe Gaussian sampler into implementations that currently use NumPy noise and re-running Delta-Siege would show whether their leaks disappear.
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 presents dpmm, an open-source Python library for differentially private synthetic tabular data generation that implements three marginal models (PrivBayes+PGM, MST, and AIM). The authors claim end-to-end DP guarantees by controlling data domain extraction, DP discretization (PrivTree or uniform), and floating-point-safe Gaussian sampling, along with enhanced functionality such as mixed data support, conditional generation, public pretraining, structural zeros, model size control, and serialization. They evaluate utility on the Wine dataset against private-pgm, OpenDP, and synthcity, and report DP auditing experiments using AuditSynth and Delta-Siege, concluding that dpmm is the only library with state-of-the-art DP synthetic data models that does not exhibit known privacy-related vulnerabilities.

Significance. If the claims hold, dpmm is a practically valuable contribution: it is pip-installable, Apache-2.0 licensed, includes three widely used marginal models with uniform APIs, and directly addresses known DP pitfalls such as direct domain extraction, non-DP preprocessing, fixed random states, and floating-point precision attacks. The paper's empirical work compares against external libraries and uses external auditing attacks, so the core utility and vulnerability claims are not circular. At the same time, the significance is currently tempered by the narrowness of the utility evaluation (one dataset, no confidence intervals) and by the conditionality of the end-to-end DP guarantee, which rests on an unenforced user assumption about how the data domain was obtained.

major comments (3)
  1. [Section 2, 'Data Domain'; Section 3, 'DP Auditing'] The end-to-end DP guarantee depends on the assumption that a user-supplied data domain was not derived from the private training data, but the library does not enforce or even verify this condition. Section 2 states that the domain 'can be provided as input' or extracted with 0.5 epsilon_proc, and Section 3 explicitly says the audits assume 'the data domain is specified.' If a practitioner computes column minima/maxima or category lists from the private data and passes them to fit() without spending the preprocessing budget, the pipeline is not DP, yet the abstract and Table 1 advertise unqualified 'end-to-end DP guarantees' and 'Data Domain ✓'. This is a load-bearing point: the headline privacy claim is conditional in a way that the API cannot detect and the included audits (Fig. 3) do not stress-test. The authors should either (a) implement a programmatic safeguard (e.g., a required flag certifying that the domain is public, or a mode that forces DP domain extraction), or (b) consistently qualify the privacy claim as conditional on a public domain in the abstract, introduction, and Table 1.
  2. [Section 3, 'Utility'] The utility claims '1.5% higher than private-pgm' and '147% higher than OpenDP/synthcity' rest entirely on a single dataset (Wine) with no reported variance or confidence intervals. The text says scores are averaged over ten trained instances and ten synthetic datasets per instance, but Fig. 1 shows only point means; there is no error bars, standard deviation, or per-dataset breakdown. The abstract and contributions generalize these numbers as if they establish a library-level utility advantage. Since the central contribution includes 'higher utility than previous implementations,' the evaluation should either be expanded to multiple datasets, or the claims should be explicitly scoped to Wine with uncertainty reported. At minimum, the paper should state that the 1.5% and 147% figures are single-dataset observations, not general performance guarantees.
  3. [Section 2, Table 1; Section 3, 'DP Auditing'] The headline statement that dpmm is 'the only library containing state-of-the-art DP synthetic data models and does not exhibit (known) privacy-related vulnerabilities' is stronger than the evidence supports. The paper itself concedes that 'these audits cannot guarantee the absence of all bugs,' and the audits are run under the safe-domain assumption discussed above. Moreover, Table 1 marks OpenDP as only partially meeting the data-domain criterion and synthcity as not meeting it, but the comparison conflates 'library does not provide the option to pass a domain' with 'library is vulnerable': OpenDP's domain extraction for continuous data is DP, and its failure in AuditSynth (Fig. 3a) is attributed to not supporting discrete domains, not to a broken mechanism. The claim should be rephrased to: 'among the libraries tested, and under the stated assumptions, dpmm is the only one that both supports all three models and passes the two auditing attacks.'
minor comments (5)
  1. [Section 3, 'Utility'] The axis label in Fig. 1(b) reads 'Mean runtime (seconds; log scale)', which is clear, but the text does not report whether the runtime comparisons used identical hardware and software environments; a sentence specifying the machine and library versions would aid reproducibility.
  2. [Table 1] The footnotes '∼1', '∼2', and '∼3' are ambiguous. A reader cannot tell from the table alone what exactly is only partially satisfied; the footnotes should be expanded into a short legend or the table should be revised to state the limitation in each cell.
  3. [Section 2, 'Data Preprocessing'] The sentence 'The former is a DP tree-based method using 0.5 epsilon_proc' should specify whether the entire 0.5 epsilon_proc is consumed by PrivTree or whether this is in addition to the domain extraction budget; Section 3 says epsilon_proc = 0.1, but the split between domain extraction and preprocessing is not stated numerically.
  4. [References] Reference [11] cites 'Dheeru Dua and Casey Graff' for the UCI Machine Learning Repository, but the current repository is typically cited with the full UCI team; please update the citation to the canonical form used by the UCI ML Repository.
  5. [Section 3, 'DP Auditing'] The paragraph describing AuditSynth says '1,000 models per implementation (ϵ = 1, δ = 10−3)' but Fig. 3a does not show the δ value; adding δ to the caption would avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: dpmm's utility and privacy claims rest on external benchmarks and external auditing attacks; the self-citations used for design choices are corroborated by independent references and by the paper's own experiments.

full rationale

This is a systems/engineering paper, not a mathematical derivation. The central utility claims (Fig. 1) compare dpmm's implementations of PrivBayes, MST, and AIM against the external libraries private-pgm, OpenDP, and synthcity on the Wine dataset, using metrics external to the authors' outputs. The privacy claims are supported by external auditing procedures, AuditSynth [3] and Delta-Siege [26], that test the implementations against the theoretical epsilon; these audits can fail and are not constructed to force a particular outcome. The DP building blocks (domain extraction, PrivTree discretization, and floating-point-safe Gaussian sampling) are adopted from external sources [9, 49, 6], with the paper's own experiments and external references [44, 10, 26] providing corroboration. The authors' prior work is cited for design expectations, e.g., the superiority of PrivTree [18] and the relative performance of marginal models [16], but these citations are not the sole or load-bearing support: the paper also cites external work [44] and reports its own benchmark results. The domain caveat is explicitly stated in Section 2 ('Data Domain') and Section 3 ('assuming the data domain is specified'); it is a conditional guarantee rather than a hidden circular reduction, because the library does not define the guarantee in terms of its own output. No equation in the paper equates a predicted quantity to a fitted input, and no 'uniqueness theorem' from the authors' prior work is invoked. The minor self-citations in the motivation and design discussion justify a low nonzero score, but they do not make the central claims circular.

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

The central claim rests on standard DP composition, on the correctness of cited mechanisms, and on a single-dataset empirical evaluation. No new entities are introduced.

free parameters (3)
  • epsilon = 1
    Privacy budget for model training in experiments; not fit to data, but the reported utility numbers are conditional on this choice.
  • delta = 1e-5 (utility), 1e-3 (audits)
    Privacy parameter chosen per experiment; utility and audit results depend on this setting.
  • epsilon_proc = 0.1
    Privacy budget allocated to preprocessing (domain extraction and discretization) in experiments; utility results depend on this division of budget.
assumptions (3)
  • standard math Standard differential privacy composition applies when combining preprocessing budget (0.5 epsilon_proc) with model budget (epsilon).
    Relied on for the end-to-end DP claim in Section 2, DP Guarantees.
  • domain assumption The cited mechanisms (Gaussian mechanism, PrivTree, Desfontaines min/max extraction) satisfy their published DP guarantees.
    The paper adopts these mechanisms as black boxes from prior literature and assumes their correctness.
  • ad hoc to paper The Wine dataset and default hyperparameters are representative for evaluating utility.
    The general utility improvement claim is based on this single evaluation setup in Section 3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of dpmm: Differentially Private Marginal Models, a Library for Synthetic Tabular Data Generation." pith.science (2026). https://pith.science/paper/B2LEFZS5

@misc{pith2026250600322,
  author       = {Pith},
  title        = {Pith review of: dpmm: Differentially Private Marginal Models, a Library for Synthetic Tabular Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2LEFZS5}},
  note         = {Machine review of arXiv:2506.00322}
}
read the original abstract

We propose dpmm, an open-source library for synthetic data generation with Differentially Private (DP) guarantees. It includes three popular marginal models -- PrivBayes, MST, and AIM -- that achieve superior utility and offer richer functionality compared to alternative implementations. Additionally, we adopt best practices to provide end-to-end DP guarantees and address well-known DP-related vulnerabilities. Our goal is to accommodate a wide audience with easy-to-install, highly customizable, and robust model implementations. Our codebase is available from https://github.com/sassoftware/dpmm.

Figures

Figures reproduced from arXiv: 2506.00322 by the authors.

Figure 1
Figure 1. Comparison between dpmm and other libraries in terms of utility and runtime, all models are trained with default hyperparameters, ϵ = 1, and δ = 10−5 on Wine. DP Guarantees. Numerous popular DP libraries, including synthetic data generation ones, contain well/long-known DP vulnerabilities such as direct data domain extraction from the input data and floating-point (im)precision [3, 6, 19, 26] (see [PITH_FULL_IMAGE:… view at source ↗
Figure 2
Figure 2. Utility-privacy tradeoffs of dpmm on Wine. 0 1 2 3 4 Empirical emp PrivBayes MST AIM Theoretical dpmm (ours) private-pgm OpenDP synthcity 0 1 2 3 4 Empirical emp PrivBayes MST AIM (a) AuditSynth [3] 10 0 10 1 Empirical emp (log scale) NumPy Laplace NumPy Gaussian OpenDP Gaussian Underlying Mechanism PrivBayes MST AIM (b) Delta-Siege [26] [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 48 canonical work pages

  1. [1]

    Privacy preserving synthetic data release using deep learning

    Nazmiye Ceren Abay, Yan Zhou, Murat Kantarcioglu, Bhavani Thuraisingham, and Latanya Sweeney. Privacy preserving synthetic data release using deep learning. In ECML PKDD, 2018

    work page 2018

  2. [2]

    Differentially private mixture of generative neural networks

    Gergely Acs, Luca Melis, Claude Castelluccia, and Emiliano De Cristofaro. Differentially private mixture of generative neural networks. IEEE TKDE, 2018

    work page 2018

  3. [3]

    What do you want from theory alone?

    Meenatchi Sundaram Muthu Selva Annamalai, Georgi Ganev, and Emiliano De Cristofaro. “What do you want from theory alone?” experimenting with tight auditing of differentially private synthetic data generation. In USENIX Security, 2024

    work page 2024

  4. [4]

    Differentially private query release through adaptive projection

    Sergul Aydore, William Brown, Michael Kearns, Krishnaram Kenthapadi, Luca Melis, Aaron Roth, and Ankit A Siva. Differentially private query release through adaptive projection. In ICML, 2021

    work page 2021

  5. [5]

    Data synthesis via differentially private markov random fields

    Kuntai Cai, Xiaoyu Lei, Jianxin Wei, and Xiaokui Xiao. Data synthesis via differentially private markov random fields. PVLDB, 2021

    work page 2021

  6. [6]

    Widespread underestimation of sensitivity in differentially private libraries and how to fix it

    Sílvia Casacuberta, Michael Shoemate, Salil Vadhan, and Connor Wagaman. Widespread underestimation of sensitivity in differentially private libraries and how to fix it. In ACM CCS, 2022

    work page 2022

  7. [7]

    Ron-gauss: Enhancing utility in non-interactive private data release

    Thee Chanyaswad, Changchang Liu, and Prateek Mittal. Ron-gauss: Enhancing utility in non-interactive private data release. PoPETs, 2019

    work page 2019

  8. [8]

    Synthetic Data: Methods, Use Cases, and Risks

    Emiliano De Cristofaro. Synthetic Data: Methods, Use Cases, and Risks. IEEE S&P Magazine, 2024

    work page 2024

Show all 52 references
  1. [9]

    Lowering the cost of anonymization

    Damien Desfontaines. Lowering the cost of anonymization . PhD thesis, ETH Zurich, 2020

    2020

  2. [10]

    How to Break, Then Fix, Differential Privacy on Finite Computers

    Damien Desfontaines and Samuel Haney. How to Break, Then Fix, Differential Privacy on Finite Computers. In USENIX PEPR, 2023

    2023

  3. [11]

    UCI Machine Learning Repository

    Dheeru Dua and Casey Graff. UCI Machine Learning Repository. https://archive.ics.uci.edu/ datasets, 2017

    2017

  4. [12]

    The algorithmic foundations of differential privacy

    Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. F oundations and Trends in Theoretical Computer Science, 2014

    2014

  5. [13]

    Our data, ourselves: Privacy via distributed noise generation

    Cynthia Dwork, Krishnaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In EuroCrypt, 2006

    2006

  6. [14]

    Calibrating noise to sensitivity in private data analysis

    Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In TCC, 2006

    2006

  7. [15]

    Using Synthetic Data in Financial Services

    FCA. Using Synthetic Data in Financial Services. https://www.fca.org.uk/publication/ corporate/report-using-synthetic-data-in-financial-services.pdf , 2024

    2024

  8. [16]

    Graphical vs

    Georgi Ganev, Kai Xu, and Emiliano De Cristofaro. Graphical vs. Deep Generative Models: Measuring the Impact of Differentially Private Mechanisms and Budgets on Utility. In ACM CCS, 2024

    2024

  9. [17]

    The Elusive Pursuit of Reproducing PATE-GAN: Benchmarking, Auditing, Debugging

    Georgi Ganev, Meenatchi Sundaram Muthu Selva Annamalai, and Emiliano De Cristofaro. The Elusive Pursuit of Reproducing PATE-GAN: Benchmarking, Auditing, Debugging. TMLR, 2025

    2025

  10. [18]

    The Importance of Being Discrete: Measuring the Impact of Discretization in End-to-End Differentially Private Synthetic Data

    Georgi Ganev, Meenatchi Sundaram Muthu Selva Annamalai, Sofiane Mahiou, and Emiliano De Cristofaro. The Importance of Being Discrete: Measuring the Impact of Discretization in End-to-End Differentially Private Synthetic Data. arXiv:2504.06923, 2025

    2025

  11. [19]

    Precision-based attacks and interval refining: how to break, then fix, differential privacy on finite computers

    Samuel Haney, Damien Desfontaines, Luke Hartman, Ruchit Shrestha, and Michael Hay. Precision-based attacks and interval refining: how to break, then fix, differential privacy on finite computers. In TPDP, 2022

    2022

  12. [20]

    Differentially Private Release of Israel’s National Registry of Live Births

    Shlomi Hod and Ran Canetti. Differentially Private Release of Israel’s National Registry of Live Births. In IEEE S&P, 2025

    2025

  13. [21]

    SoK: Privacy-Preserving Data Synthesis

    Yuzheng Hu, Fan Wu, Qinbin Li, Yunhui Long, Gonzalo Munilla Garrido, Chang Ge, Bolin Ding, David Forsyth, Bo Li, and Dawn Song. SoK: Privacy-Preserving Data Synthesis. In IEEE S&P, 2024

    2024

  14. [22]

    PATE-GAN: generating synthetic data with differential privacy guarantees

    James Jordon, Jinsung Yoon, and Mihaela Van Der Schaar. PATE-GAN: generating synthetic data with differential privacy guarantees. In ICLR, 2018

    2018

  15. [23]

    Synthetic Data–what, why and how? arXiv:2205.03257, 2022

    James Jordon, Lukasz Szpruch, Florimond Houssiau, Mirko Bottarelli, Giovanni Cherubin, Carsten Maple, Samuel N Cohen, and Adrian Weller. Synthetic Data–what, why and how? arXiv:2205.03257, 2022

    2022 arXiv

  16. [24]

    Differentially private synthesization of multi-dimensional data using copula functions

    Haoran Li, Li Xiong, and Xiaoqian Jiang. Differentially private synthesization of multi-dimensional data using copula functions. In EDBT, 2014

    2014

  17. [25]

    Iterative methods for private synthetic data: Unifying framework and new methods

    Terrance Liu, Giuseppe Vietri, and Steven Z Wu. Iterative methods for private synthetic data: Unifying framework and new methods. NeurIPS, 2021

    2021

  18. [26]

    Dimitrov, and Martin Vechev

    Johan Lokna, Anouk Paradis, Dimitar I. Dimitrov, and Martin Vechev. Group and Attack: Auditing Differential Privacy. In ACM CCS, 2023

    2023

  19. [27]

    Gunter, and Bo Li

    Yunhui Long, Boxin Wang, Zhuolin Yang, Bhavya Kailkhura, Aston Zhang, Carl A. Gunter, and Bo Li. G-PATE: Scalable differentially private data generator via private aggregation of teacher discriminators. In NeurIPS, 2021. 5

    2021

  20. [28]

    dpart: Differentially private autoregressive tabular, a general framework for synthetic data generation

    Sofiane Mahiou, Kai Xu, and Georgi Ganev. dpart: Differentially private autoregressive tabular, a general framework for synthetic data generation. In TPDP, 2022

    2022

  21. [29]

    private-pgm

    Ryan McKenna. private-pgm. https://github.com/ryan112358/private-pgm, 2019

    2019

  22. [30]

    A simple recipe for private synthetic data generation

    Ryan McKenna and Terrance Liu. A simple recipe for private synthetic data generation. DifferentialPri- vacy.org, 2022. https://differentialprivacy.org/synth-data-1/

    2022

  23. [31]

    Graphical-model based estimation and inference for differential privacy

    Ryan McKenna, Daniel Sheldon, and Gerome Miklau. Graphical-model based estimation and inference for differential privacy. In ICML, 2019

    2019

  24. [32]

    Winning the NIST Contest: A scalable and general approach to differentially private synthetic data

    Ryan McKenna, Gerome Miklau, and Daniel Sheldon. Winning the NIST Contest: A scalable and general approach to differentially private synthetic data. JPC, 2021

    2021

  25. [33]

    AIM: an adaptive and iterative mechanism for differentially private synthetic data

    Ryan McKenna, Brett Mullins, Daniel Sheldon, and Gerome Miklau. AIM: an adaptive and iterative mechanism for differentially private synthetic data. PVLDB, 2022

    2022

  26. [34]

    Mechanism design via differential privacy

    Frank McSherry and Kunal Talwar. Mechanism design via differential privacy. In FOCS, 2007

    2007

  27. [35]

    IOM and Microsoft release first-ever differentially private synthetic dataset to counter human trafficking

    Microsoft. IOM and Microsoft release first-ever differentially private synthetic dataset to counter human trafficking. https://www.microsoft.com/en-us/research/blog/iom-and-microsoft- release-first-ever-differentially-private-synthetic-dataset-to-counter-human- trafficking/, 2022

    2022

  28. [36]

    2020 Census Data Products: Data Needs and Privacy Considerations: Proceedings of a Workshop

    NASEM. 2020 Census Data Products: Data Needs and Privacy Considerations: Proceedings of a Workshop. The National Academies Press, 2020

    2020

  29. [37]

    Adversary instantiation: Lower bounds for differentially private machine learning

    Milad Nasr, Shuang Songi, Abhradeep Thakurta, Nicolas Papernot, and Nicholas Carlin. Adversary instantiation: Lower bounds for differentially private machine learning. In IEEE S&P, 2021

    2021

  30. [38]

    Tight Auditing of Differentially Private Machine Learning

    Milad Nasr, Jamie Hayes, Thomas Steinke, Borja Balle, Florian Tramèr, Matthew Jagielski, Nicholas Carlini, and Andreas Terzis. Tight Auditing of Differentially Private Machine Learning. In USENIX Security, 2023

    2023

  31. [39]

    2018 Differential privacy synthetic data challenge

    NIST. 2018 Differential privacy synthetic data challenge. https://www.nist.gov/ctl/pscr/open- innovation-prize-challenges/past-prize-challenges/2018-differential-privacy- synthetic, 2018

    2018

  32. [40]

    Synthesising the linked 2011 Census and deaths dataset while preserving its confidential- ity

    ONS. Synthesising the linked 2011 Census and deaths dataset while preserving its confidential- ity. https://datasciencecampus.ons.gov.uk/synthesising-the-linked-2011-census-and- deaths-dataset-while-preserving-its-confidentiality/ , 2023

    2011

  33. [41]

    SmartNoise SDK: Tools for Differential Privacy on Tabular Data

    OpenDP. SmartNoise SDK: Tools for Differential Privacy on Tabular Data. https://github.com/ opendp/smartnoise-sdk, 2021

    2021

  34. [42]

    DataSynthesizer: Privacy-Preserving Synthetic Datasets

    Haoyue Ping, Julia Stoyanovich, and Bill Howe. DataSynthesizer: Privacy-Preserving Synthetic Datasets. In SSDBM, 2017

    2017

  35. [43]

    Synthcity: a benchmark framework for diverse use cases of tabular synthetic data

    Zhaozhi Qian, Rob Davis, and Mihaela van der Schaar. Synthcity: a benchmark framework for diverse use cases of tabular synthetic data. In NeurIPS Datasets and Benchmarks Track, 2023. https://github. com/vanderschaarlab/synthcity

    2023

  36. [44]

    Benchmarking differentially private synthetic data generation algorithms

    Yuchao Tao, Ryan McKenna, Michael Hay, Ashwin Machanavajjhala, and Gerome Miklau. Benchmarking differentially private synthetic data generation algorithms. In PPAI, 2022

    2022

  37. [45]

    Synthetic data to test the effectiveness of a vulnerable person’s detection system in financial ser- vices

    UK ICO. Synthetic data to test the effectiveness of a vulnerable person’s detection system in financial ser- vices. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data- sharing/privacy-enhancing-technologies/case-studies/synthetic-data-to-test-the- effecti...

    2023

  38. [46]

    New oracle-efficient algorithms for private synthetic data release

    Giuseppe Vietri, Grace Tian, Mark Bun, Thomas Steinke, and Steven Wu. New oracle-efficient algorithms for private synthetic data release. In ICML, 2020

    2020

  39. [47]

    Private synthetic data for multitask learning and marginal queries

    Giuseppe Vietri, Cedric Archambeau, Sergul Aydore, William Brown, Michael Kearns, Aaron Roth, Ankit Siva, Shuai Tang, and Steven Z Wu. Private synthetic data for multitask learning and marginal queries. NeurIPS, 2022

    2022

  40. [48]

    Differentially private generative adversarial network

    Liyang Xie, Kaixiang Lin, Shu Wang, Fei Wang, and Jiayu Zhou. Differentially private generative adversarial network. arXiv:1802.06739, 2018

    2018 arXiv

  41. [49]

    Privtree: A differentially private algorithm for hierarchical decompositions

    Jun Zhang, Xiaokui Xiao, and Xing Xie. Privtree: A differentially private algorithm for hierarchical decompositions. In SIGMOD, 2016

    2016

  42. [50]

    Procopiuc, Divesh Srivastava, and Xiaokui Xiao

    Jun Zhang, Graham Cormode, Cecilia M. Procopiuc, Divesh Srivastava, and Xiaokui Xiao. PrivBayes: Private Data Release via Bayesian Networks. ACM TDS, 2017

    2017

  43. [51]

    Differentially private releasing via deep generative model (technical report)

    Xinyang Zhang, Shouling Ji, and Ting Wang. Differentially private releasing via deep generative model (technical report). arXiv:1801.01594, 2018

    2018 arXiv

  44. [52]

    PrivSyn: Differentially Private Data Synthesis

    Zhikun Zhang, Tianhao Wang, Jean Honorio, Ninghui Li, Michael Backes, Shibo He, Jiming Chen, and Yang Zhang. PrivSyn: Differentially Private Data Synthesis. In USENIX Security, 2021. 6

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.