Pith. sign in

REVIEW 3 major objections 5 minor 72 references

This paper shows that the optimal DP noise for histogram-based synthetic data is a floor-raised copy of the public distribution, and that using it as the noise source improves the privacy-utility trade-off.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 16:19 UTC pith:FZTBS7FY

load-bearing objection A credible theory result—the asymptotically optimal domain-aware noise is a floor-raised public distribution—with a practical gap between the solved LP and the deployed approximation, plus hand-set gamma in two of the main experiments. the 3 major comments →

arxiv 2607.18021 v1 pith:FZTBS7FY submitted 2026-07-20 cs.IT cs.CRmath.IT

Optimal Domain-Aware Privacy Mechanisms for Synthetic Data Generation

Sajani Vithana , Sangwon Jung , Haoyang Hu , Viveck R. Cadambe , Flavio P. Calmon , Haewon Jeong This is my paper
classification cs.IT cs.CRmath.IT MSC 68P2794A17
keywords differential privacysynthetic datapublic datadomain-aware mechanismhistogramlinear mixingfloor-raisingprivacy-utility trade-off
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's aim is to prove that public data can be used not merely to pre-train models or align outputs, but to design the very noise distribution that provides differential privacy. For histogram-based synthetic data generation, it studies mechanisms that release m samples from the mixing Psyn=(1-beta) private histogram + beta noise, and shows the asymptotically optimal noise is a floor-raised version of the public histogram. The result is constructive: PubMix is a drop-in replacement for the Gaussian or Laplace noise used in histogram-sampling synthesis pipelines, and the experiments show it consistently improves synthetic-data utility at the same privacy budget. If the proof holds, a previously heuristic benefit of public data becomes a precise mechanism-design rule.

Core claim

The paper establishes that inside the class of linear mixing mechanisms, the privacy constraint depends on the noise distribution only through its two smallest masses. Theorem 4.2 then reduces the worst-case utility optimization to a one-dimensional search over a floor level t, solving a linear program for the radius f(t); the optimal noise Pnse is a floor-raised version of the public distribution Pbar_pub, and the optimal mixing weight is beta*=beta_min(t*,t*). Without public data, the optimal noise is uniform, so the public ball B_{gamma,Ppub} is what redirects the perturbation toward realistic directions. This is the sense in which domain-aware DP noise outperforms domain-agnostic Gaussia

What carries the argument

The central mechanism is linear mixing, Psyn=(1-beta)Pbar_prv+beta Pnse, where Pnse is the noise distribution and beta the privacy weight. Because neighboring private histograms differ in at most two coordinates, the DP constraint tightens to a condition on the two smallest masses of Pnse, Nmin1 and Nmin2; beta_min(Nmin1,Nmin2) is the smallest mixing weight that satisfies (epsilon,delta)-DP. Floor-raising (raise bins below threshold t to t, take the excess from bins above t) is the operation that shapes Pnse: it raises Nmin1 to lower beta_min while staying near Pbar_pub to keep the worst-case TV radius small. The exact optimum is characterized by the LP in (12); PubMix uses two computable fl

Load-bearing premise

The load-bearing premise is that the actual private histogram lies inside the gamma-TV ball centered on the public histogram, with gamma chosen by hand in the experiments rather than validated; if that ball misses the private data, the utility gain over Gaussian noise is not covered by the theory.

What would settle it

Two concrete checks: (1) brute-force solve the original min-max problem (3) for a finite alphabet (say d=4) and compare its value to min_{0<=t<=1/d} beta_min(t,t) f(t) from Theorem 4.2—a gap that does not vanish as d grows would refute the asymptotic optimality claim; (2) measure the empirical TV distance between the actual private histogram and the public histogram in a deployment where gamma was set by hand (0.9 and 0.1), and check whether the private histogram is inside B_{gamma,Ppub}.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • At a fixed privacy budget, replacing domain-agnostic noise with PubMix improves the TV distance between private and synthetic distributions; in the reported experiments the gain is largest at the strictest privacy level (epsilon=1).
  • Because beta_min depends only on the two smallest masses of Pnse, the design problem decouples: only the tail of the noise distribution controls the privacy cost, which is why floor-raising is the right operation.
  • Without any public data, the framework recovers uniform mixing as optimal; public data is exactly what shrinks the worst-case region from the whole simplex to B_{gamma,Ppub}.
  • The same floor-raised construction applies to general discrete estimators with bounded sensitivity s, with beta increased to compensate for losing the two-point sparsity of histograms.
  • As the alphabet size d grows, the relaxed upper-bound problem converges to the original problem, so the LP solution is asymptotically optimal.

Where Pith is reading between the lines

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

  • (Editorial inference) The same recipe could improve private prediction systems that already mix next-token distributions toward a public distribution: instead of fixing the public distribution as noise, optimize a floor-raised version of it, which should lower the required mixing weight.
  • (Editorial inference) Gamma is the real pivot of the method: a principled calibration would set gamma from the empirical quantile of pairwise fold distances and then verify coverage on held-out data; the reported experiments set it by hand (0.9 and 0.1), so gamma's value should be treated as a hyperparameter until such a procedure exists.
  • (Editorial inference) The asymptotic optimality in d suggests a finite-d refinement: explicitly optimizing Nmin1 and Nmin2, rather than replacing Nmin2 by Nmin1, would shrink the suboptimality gap that the practical floor-raising approximations are admitted to have.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a public-data-aware differential privacy mechanism for histogram-based synthetic data generation. It models the synthetic distribution as a linear mix Psyn = (1-beta) Pbar_prv + beta Pnse and formulates a minimax problem over a gamma-TV ball around the public distribution Pbar_pub. The main theoretical results are: (i) Theorem 4.1 gives the minimum mixing weight beta_min for a fixed noise distribution, showing it depends only on the two smallest masses of Pnse; (ii) Theorem 4.2 characterizes an asymptotically optimal noise distribution as a floor-raised version of the public distribution, obtained by a linear program; (iii) without public data the optimal noise is uniform (Corollary 4.6). The practical mechanism PubMix uses a MIFR/water-filling approximation of the LP solution. Experiments on PUMS, Person Activity, and Yelp report improved downstream utility compared to domain-agnostic baselines when PubMix is used as a drop-in component in private-evolution pipelines.

Significance. If the main theorem is correct, the paper makes a useful conceptual contribution: it shows how public data can be incorporated directly into the noise-design step of DP mechanisms, not only in pre/post-processing. The reduction of the DP condition to a hockey-stick divergence over two changed coordinates is clean, and the proofs in Appendices C-E are detailed and plausible. The LP characterization in Theorem 4.2 is a concrete, falsifiable design rule. The empirical evaluation is also modular and compares against a strong recent baseline (TabPE, PE). However, the practical claims are conditional in two load-bearing ways: the deployed mechanism uses an admitted sub-optimal approximation (Remark 4.5), and the gamma-ball assumption is not validated in the main experiments. These gaps do not invalidate the theory, but they mean the paper's headline 'optimal domain-aware mechanism' and its utility improvements are not directly supported by the presented analysis.

major comments (3)
  1. [Remark 4.5, Algorithm 1, App. F] The mechanism evaluated in Section 5 is not the optimizer of Theorem 4.2. Remark 4.5 states that the MIFR/WF approximation is 'generally sub-optimal', and App. F provides only an upper bound on the objective, not a bound on the suboptimality gap relative to the LP optimum. Since Algorithm 1 instantiates this approximation, the experiments demonstrate utility of a mechanism whose performance relative to the central optimality theorem is unquantified. Moreover, the LP in (12) has 2^d subset constraints, so it is not obvious that the exact optimum is even computable for the d values used in the text experiments. Please either solve the LP (or provide a polynomial-time method) or quantify the gap between MIFR/WF and the LP optimum.
  2. [Sec. 5, Apps. I.1/I.2, Eq. (3), footnote 1] The utility guarantee in Theorem 4.2 is a minimax guarantee over the gamma-TV ball B_{gamma,Pbar_pub}. The paper's own footnote 1 concedes that if the true private distribution lies outside this ball, only a weaker, unquantified utility guarantee holds. In the Person Activity experiment gamma is simply set to 0.9 (App. I.1) and in the Yelp experiment to 0.1 (App. I.2). The pairwise-TV estimation procedure described in Sec. 5 is not used there, and no coverage or misestimation analysis is provided. Without validating that the actual private histograms lie in the chosen ball, the reported utility improvements are not backed by the theory. Please add a coverage analysis, a sensitivity analysis over gamma, or substantially qualify the practical claims.
  3. [Theorem 4.2 and Sec. 3.1 (Eq. (5))] The asymptotic optimality statement is made as d -> infinity with gamma constant, but the paper does not give finite-d bounds on the gap between the original problem in (3) and the upper-bound problem in (9). The continuity argument in App. D.1 is plausible, but the result is asymptotic; the experiments use finite, sometimes modest, d. A concrete finite-d bound or an experimental check of the gap would strengthen the claim that the floor-raised solution is the right design rule in practice.
minor comments (5)
  1. [Table 1 and Sec. 5.2.1] The text calls the method TABPE+PUBMIX, but Table 1 labels it PUBMIX. Clarify the name used in the table.
  2. [App. I.2] The text says 'We report accuracy for the category task and MAE for the star regression task,' but Tables 2 and 3 report 'Rating RMSE'. Align terminology.
  3. [Eq. (12)] The constraint is written as '1 T_d Pnse = 1'; use a cleaner notation such as sum_i Pnse[i] = 1. Also define Pnse(S) before using it in the LP.
  4. [References] Several references contain typos, e.g., 'availabe at:' in Ghalebikesabi et al., Frigerio et al., Pfitzner & Arnrich, and Dockhorn et al. Please correct.
  5. [Fig. 2] The axes and curves are informative, but the caption does not explain why beta_min is flat for small Nmin or the role of m. A short explanation would help.

Circularity Check

0 steps flagged

No significant circularity: Theorem 4.2 solves the stated minimax problem; the gamma-ball is an explicit assumption, not a disguised output.

full rationale

The central derivation chain is self-contained rather than circular. Theorem 4.1 computes the minimal mixing weight beta_min from the hockey-stick divergence of the m-fold product distributions, with the two-point sparsity of neighboring histograms used explicitly. Theorem 4.2 does not assume the floor-raised form; it derives it by rewriting the minimax objective, introducing t = min_i Pnse[i], and solving the LP in (12), whose constraints r + Pnse(S) >= min{Pbar_pub(S)+gamma, 1} come from Lemma E.1's exact characterization of the worst-case TV over B_{gamma,Pbar_pub}. The optimal floor-raised Pnse is therefore a consequence of the optimization, not an input. The admissible set B_{gamma,Pbar_pub} is an assumption about the domain (private distributions lie near the public distribution); the paper's footnote 1 explicitly concedes that outside this ball only a weaker, unquantified utility guarantee holds. That is a conditional guarantee and a validation concern, not circularity. The gamma values used in the experiments (0.9 and 0.1 in Apps. I.1/I.2) are hand-set hyperparameters; if they are misspecified, the theoretical utility guarantee may not apply, but this is a correctness/robustness issue, not an equation reducing a prediction to a fit. Remark 4.5 honestly states that the MIFR/WF approximation is generally sub-optimal, so no fitted quantity is being relabeled as optimal. No load-bearing self-citations appear: the only invoked external results (Gallager's floor-raising, Sason & Verdu's data-processing inequality for hockey-stick divergence) are standard and not from the authors' prior work. The empirical comparisons against external baselines (AIM, GSD, GEM, PrivSyn, TabPE, PE) are self-contained. Therefore the paper's main theoretical result is not equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No new physical or ontological entities are introduced. The free parameters are gamma and the floor-raising thresholds; both are tuned rather than derived. The domain assumptions are the closeness of private/public distributions and the normalized-histogram adjacency structure. The mathematical axioms are standard f-divergence and continuity properties. The practical approximation of Theorem 4.2 by MIFR/WF is an unproven design choice, not a fitted parameter per se.

free parameters (2)
  • proximity parameter gamma = PUMS: average pairwise state TV; Person Activity: 0.9; Yelp: 0.1
    Defines the admissible ball B_{gamma,Pbar_pub} and enters every subsequent choice of Pnse and beta (Sec. 2). It is estimated from public folds or set by hand, not derived from a formal coverage guarantee.
  • floor-raising thresholds t1/t2 (and LP threshold t*) = optimized numerically; no closed form or exact values reported
    The optimum t in Theorem 4.2 and Remark 4.5 is found by scanning/minimizing beta_min(t,t) f(t). The paper does not specify a scalable algorithm or exact resulting values, so this acts as a tuned design parameter.
axioms (5)
  • domain assumption Neighboring private histograms differ in exactly two coordinates by +/-1/n
    Used to reduce the privacy analysis to the two-point sparse structure in Lemma C.1; holds for normalized histograms under the replacement adjacency used in the paper.
  • domain assumption The actual private distribution lies in B_{gamma,Pbar_pub}
    The optimization in (3) minimizes worst-case TV over this ball. If violated, privacy still holds but the utility guarantee is void (footnote 1, Sec. 2).
  • domain assumption Dprv and Dpub are i.i.d. samples from the same domain
    The public distribution guides the noise; if the public dataset is not from the same domain, floor-raising Ppub is not a meaningful target for preserving Pprv structure.
  • standard math Hockey-stick divergence satisfies the data processing inequality
    Used in Lemma C.2 and Claim D.2 to establish monotonicity of the privacy loss and feasibility of larger beta.
  • standard math beta_min is continuous and non-increasing in Nmin1, Nmin2
    Proved in App. C-D; required for the asymptotic optimality argument that replaces beta_min(N1,N2) with beta_min(N1,N1).

pith-pipeline@v1.3.0-alltime-deepseek · 39719 in / 12236 out tokens · 140867 ms · 2026-08-01T16:19:06.247089+00:00 · methodology

0 comments
read the original abstract

Differential privacy (DP) imposes fundamental trade-offs between privacy and statistical fidelity in synthetic data generation. While access to public data has been shown to improve these trade-offs empirically, existing approaches use public data only indirectly, through pre-processing (e.g., using pre-trained generative models) or post-processing steps (e.g., matching target statistics estimated from public datasets), while relying on domain-agnostic DP mechanisms. In this work, we lay the theoretical framework to study the principled incorporation of public data into DP mechanisms themselves. We consider normalized histograms as distribution estimators and characterize the asymptotically optimal domain-aware privacy mechanism within a specific class of DP mechanisms. We introduce PubMix, a public-data-aware DP mechanism that can be used in histogram-based data synthesis pipelines. Our experiments demonstrate that PubMix significantly improves synthetic data generation quality compared to domain-agnostic privacy mechanisms.

Figures

Figures reproduced from arXiv: 2607.18021 by Flavio P. Calmon, Haewon Jeong, Haoyang Hu, Sajani Vithana, Sangwon Jung, Viveck R. Cadambe.

Figure 1
Figure 1. Figure 1: Comparison of PUBMIX, LINMIX, and the Gaussian Mechanism for differentially private histograms. (a)–(c) provide a geometric illustration of the perturbations in the probability simplex for (a) PUBMIX, (b) LINMIX, and (c) the Gaussian mechanism. (d) shows the total variation distance between (P¯prv, Psyn) as a function of the number of synthetic samples m, for PUBMIX, LINMIX, and the Gaussian mechanism for … view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of minimum mixing weight βmin for differ￾ent numbers of synthetic samples m and minimum noise compo￾nents Nmin with n = 1200, d = 1000, ε = 5 and δ = 1/n. (a) Original: P (b) Water-filled(P, t) (c) MIFR(P, t) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two instances of floor-raising: (a) original PMF, (b) Water-filling: raises smaller masses to t and removes the added mass uniformly from larger masses (Def. 4.3), (c) Maximum Invari￾ant Floor Raising (MIFR): raises smaller masses to t and removes the added mass Madd from larger ones to maximize the unchanged probability mass with respect to the original (Def. 4.4). tion to (3) when no public data is avail… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of the total variation (TV) distance between P¯prv and the empirical distribution of the synthetic dataset Dsyn for n = 57,717, ε = 5, and δ = 1/n on the PUMS dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy of a regression model trained on synthetic data by PUBMIX, LINMIX, and the Gaussian mechanism, measured relative to a model trained on real data from the PUMS dataset [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The two cases in Remark 4.5 correspond to two sets of potential noise distributions: 1) a worst-case P¯prv in (3) that is on the same line as Pnse and P¯pub; 2) a worst-case P¯prv that is not on the same line. The Key insight from Remark 4.5 is that the approximation to the optimal noise distribution Pˆ∗ nse is always a floor-raised ver￾sion of P¯ pub, where both MIFR(P¯ pub, t1) and WF(P¯ pub, t2) raise s… view at source ↗
Figure 7
Figure 7. Figure 7: Histograms over candidate P¯pub distributions, showing the difference between the minimax TV distance g(Pnse) = minβ≥βmin maxP¯prv∈Bγ,P¯ pub TV(P¯prv, Psyn) considering the optimal P ∗ nse from Theorem 4.2 and the approximation Pˆ∗ nse in Re￾mark 4.5. F.1. Proof of Lemma F.1 Proof: For a fixed Pnse ∈ A with mini P [i] nse = Nmin1 , g(Pnse) = min β∈[0,1] β≥βmin(Nmin1 ,Nmin1 ) max P¯prv∈Bγ,P¯ pub TV(P¯ prv, … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

72 extracted references · 7 linked inside Pith

  1. [1]

    IEEE Transactions on Information Theory , volume=

    f -divergence Inequalities , author=. IEEE Transactions on Information Theory , volume=. 2016 , publisher=

    2016

  2. [2]

    Ponomareva and Z

    N. Ponomareva and Z. Xu and H. B. McMahan and P. Kairouz and L. Rosenblatt and V. Cohen-Addad and C. Guzmán and R. McKenna and G. Andrew and A. Bie and D. Yu and A. Kurakin and M. Zadimoghaddam and S. Vassilvitskii and A. Terzis , title=. arXiv preprint arXiv:2512.03238 , year=

    Pith/arXiv arXiv

  3. [3]

    Proceedings of The 27th International Conference on Artificial Intelligence and Statistics , pages =

    Joint Selection: Adaptively Incorporating Public Information for Private Synthetic Data , author =. Proceedings of The 27th International Conference on Artificial Intelligence and Statistics , pages =. 2024 , volume =

    2024

  4. [4]

    arXiv:2306.01684 , year=

    Alexey Kurakin and Natalia Ponomareva and Umar Syed and Liam MacDermed and Andreas Terzis , title=. arXiv:2306.01684 , year=

    Pith/arXiv arXiv

  5. [5]

    Frigerio and A

    L. Frigerio and A. Oliveira and L. Gomez and P. Duverger , title=. ICT Systems Security and Privacy Protection , year=

  6. [6]

    Xie and K

    L. Xie and K. Lin and S. Wang and F. Wang and J. Zhou , title=. , journal =

  7. [7]

    Abacha and S.G

    F.Z. Abacha and S.G. Teo and L.C. Cordeiro and M.A. Mustafa , title=. Federated Learning in the Age of Foundation Models - FL 2024 International Workshops , year=

    2024

  8. [8]

    36th International Conference on Algorithmic Learning Theory , year=

    Differentially Private Multi-Sampling from Distributions , author=. 36th International Conference on Algorithmic Learning Theory , year=

  9. [9]

    Differentially Private Next-Token Prediction of Large Language Models

    Flemings, James and Razaviyayn, Meisam and Annavaram, Murali. Differentially Private Next-Token Prediction of Large Language Models. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024

    2024

  10. [10]

    Ginart and Laurens van der Maaten and James Zou and Chuan Guo , title=

    Antonio A. Ginart and Laurens van der Maaten and James Zou and Chuan Guo , title=. availabe at: arXiv:2201.00971 , year=

    Pith/arXiv arXiv

  11. [11]

    Differentially Private Synthetic Data via

    Toan Tran and Arturs Backurs and Zinan Lin and Victor Reis and Li Xiong and Sergey Yekhanin , year=. Differentially Private Synthetic Data via

  12. [12]

    Forty-second International Conference on Machine Learning , year=

    Private Federated Learning using Preference-Optimized Synthetic Data , author=. Forty-second International Conference on Machine Learning , year=

  13. [13]

    Pfitzner and B

    B. Pfitzner and B. Arnrich , title=. availabe at: arXiv:2211.11591 , year=

    Pith/arXiv arXiv

  14. [14]

    Smith and Olivia Wiles and Borja Balle , title=

    Sahra Ghalebikesabi and Leonard Berrada and Sven Gowal and Ira Ktena and Robert Stanforth and Jamie Hayes and Soham De and Samuel L. Smith and Olivia Wiles and Borja Balle , title=. availabe at: arXiv:2302.13861 , year=

    Pith/arXiv arXiv

  15. [15]

    Seemanand A

    J. Seemanand A. Slavkovic and M. Reimherr , title=. Privacy in Statistical Databases , year=

  16. [16]

    Jiang and G

    D. Jiang and G. Zhang and M. Karami and X. Chen and Y. Shao and Y. Yu , title=. , journal =

  17. [17]

    , journal =

    Yuchao Tao and Ryan McKenna and Michael Hay and Ashwin Machanavajjhala and Gerome Miklau , title=. , journal =

  18. [18]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Leveraging Public Data for Practical Private Query Release , author =. Proceedings of the 38th International Conference on Machine Learning , pages =

  19. [19]

    Cai, Kuntai and Lei, Xiaoyu and Wei, Jianxin and Xiao, Xiaokui , title =. Proc. VLDB Endow. , pages =. 2021 , issue_date =

    2021

  20. [20]

    Hay, Michael and Rastogi, Vibhor and Miklau, Gerome and Suciu, Dan , title =. Proc. VLDB Endow. , pages =. 2010 , issue_date =

    2010

  21. [21]

    2018 , booktitle =

    Cormode, Graham and Kulkarni, Tejas and Srivastava, Divesh , title =. 2018 , booktitle =

    2018

  22. [22]

    2017 , booktitle =

    Wang, Tianhao and Blocki, Jeremiah and Li, Ninghui and Jha, Somesh , title =. 2017 , booktitle =

    2017

  23. [23]

    2024 , howpublished =

    2024

  24. [24]

    2025 , url=

    Jianqing Zhang and Yang Liu and JIE FU and Yang Hua and Tianyuan Zou and Jian Cao and Qiang Yang , booktitle=. 2025 , url=

    2025

  25. [25]

    Differentially Private Synthetic Data with Private Density Estimation , year=

    Bojkovic, Nikolija and Loh, Po-Ling , booktitle=. Differentially Private Synthetic Data with Private Density Estimation , year=

  26. [26]

    2021 , booktitle =

    Raskhodnikova, Sofya and Sivakumar, Satchit and Smith, Adam and Swanberg, Marika , title =. 2021 , booktitle =

    2021

  27. [27]

    Transactions on Machine Learning Research , issn=

    Differentially Private Diffusion Models , author=. Transactions on Machine Learning Research , issn=. 2023 , note=

    2023

  28. [28]

    Proceedings of The 36th International Conference on Algorithmic Learning Theory , pages =

    Differentially Private Multi-Sampling from Distributions , author =. Proceedings of The 36th International Conference on Algorithmic Learning Theory , pages =. 2025 , editor =

    2025

  29. [29]

    Evaluating Differentially Private Synthetic Data Generation in High-Stakes Domains

    Ramesh, Krithika and Gandhi, Nupoor and Madaan, Pulkit and Bauer, Lisa and Peris, Charith and Field, Anjalie. Evaluating Differentially Private Synthetic Data Generation in High-Stakes Domains. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024

    2024

  30. [30]

    2021 , journal=

    Ryan McKenna and Gerome Miklau and Daniel Sheldon , title=. 2021 , journal=

    2021

  31. [31]

    The Twelfth International Conference on Learning Representations , year=

    Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation , author=. The Twelfth International Conference on Learning Representations , year=

  32. [32]

    Park and Y

    J. Park and Y. Choi and J. Lee , title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  33. [33]

    Yue and H

    X. Yue and H. Inan and X. Li and G. Kumar and J. McAnallen and H. Sun and D. Levitan and R. Sim , title =. ACL 2023 , year =

    2023

  34. [34]

    Dwork, Cynthia and Roth, Aaron , title =. Found. Trends Theor. Comput. Sci. , pages =. 2014 , volume =

    2014

  35. [35]

    Theory of Cryptography , year=

    Dwork, Cynthia and McSherry, Frank and Nissim, Kobbi and Smith, Adam , title=. Theory of Cryptography , year=

  36. [36]

    A Simple and Practical Algorithm for Differentially Private Data Release , year =

    Hardt, Moritz and Ligett, Katrina and Mcsherry, Frank , booktitle =. A Simple and Practical Algorithm for Differentially Private Data Release , year =

  37. [37]

    and Srivastava, Divesh and Xiao, Xiaokui , title =

    Zhang, Jun and Cormode, Graham and Procopiuc, Cecilia M. and Srivastava, Divesh and Xiao, Xiaokui , title =. 2017 , volume =

    2017

  38. [38]

    Qardaji, Wahbeh and Yang, Weining and Li, Ninghui , title =. Proc. VLDB Endow. , pages =. 2013 , volume =

    2013

  39. [39]

    , title =

    Gallager, Robert G. , title =. 1968 , isbn =

    1968

  40. [40]

    2023 , booktitle =

    Wang, Hao and Sudalairaj, Shivchander and Henning, John and Greenewald, Kristjan and Srivastava, Akash , title =. 2023 , booktitle =

    2023

  41. [41]

    Differentially Private Language Models for Secure Data Sharing

    Mattern, Justus and Jin, Zhijing and Weggenmann, Benjamin and Schoelkopf, Bernhard and Sachan, Mrinmaya. Differentially Private Language Models for Secure Data Sharing. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022

    2022

  42. [42]

    Brendan and Mironov, Ilya and Talwar, Kunal and Zhang, Li , title =

    Abadi, Martin and Chu, Andy and Goodfellow, Ian and McMahan, H. Brendan and Mironov, Ilya and Talwar, Kunal and Zhang, Li , title =. 2016 , pages =

    2016

  43. [43]

    arXiv:1802.08908 , year=

    Nicolas Papernot and Shuang Song and Ilya Mironov and Ananth Raghunathan and Kunal Talwar and Úlfar Erlingsson , title=. arXiv:1802.08908 , year=

    Pith/arXiv arXiv

  44. [44]

    Long, Yunhui and Wang, Boxin and Yang, Zhuolin and Kailkhura, Bhavya and Zhang, Aston and Gunter, Carl and Li, Bo , booktitle =

  45. [45]

    Xie and Z

    C. Xie and Z. Lin and A. Backurs and S. Gopi and D. Yu and H. Inan and H. Nori and H. Jiang and H. Zhang and Y. T. Lee and B. Li and S. Yekhanin , booktitle=. Differentially Private Synthetic Data via Foundation Model

  46. [46]

    Lin and S

    Z. Lin and S. Gopi and J. Kulkarni and H. Nori and S. Yekhanin , booktitle=. DIFFERENTIALLY PRIVATE SYNTHETIC DATA VIA FOUNDATION MODEL

  47. [47]

    Hou and A

    C. Hou and A. Shrivastava and H. Zhan, R. Conway and T. Le and A. Sagar and G. Fanti and D. Lazar , booktitle=. PrE-Text: Training Language Models on Private Federated Data in the Age of LLMs , year=

  48. [48]

    Hou and M.Y

    C. Hou and M.Y. Wang and Y. Zhu and D. Lazar and G. Fanti , booktitle=. POPri: Private Federated Learning using Preference-Optimized Synthetic Data , year=

  49. [49]

    30th USENIX Security Symposium (USENIX Security 21) , year =

    Zhikun Zhang and Tianhao Wang and Ninghui Li and Jean Honorio and Michael Backes and Shibo He and Jiming Chen and Yang Zhang , title =. 30th USENIX Security Symposium (USENIX Security 21) , year =

  50. [50]

    Graphical-model based estimation and inference for differential privacy , year=

    Ryan McKenna and Daniel Sheldon and Gerome Miklau , booktitle=. Graphical-model based estimation and inference for differential privacy , year=

  51. [51]

    Boedihardjo and T

    M. Boedihardjo and T. Strohmer and R. Vershynin , journal=. Privacy of Synthetic Data: A Statistical Framework , year=

  52. [52]

    Bojkovic and P

    N. Bojkovic and P. Loh , booktitle=. Differentially Private Synthetic Data with Private Density Estimation , year=

  53. [53]

    availabe at: arXiv preprint arXiv:2106.12118 , year=

    Ryan McKenna and Gerome Miklau and Michael Hay and Ashwin Machanavajjhala , title=. availabe at: arXiv preprint arXiv:2106.12118 , year=

    Pith/arXiv arXiv

  54. [54]

    , journal =

    Mengmeng Yang and Chi-Hung Chi and Kwok-Yan Lam and Jie Feng and Taolin Guo and Wei Ni , title=. , journal =

  55. [55]

    2022 , booktitle =

    Ye, Jiayuan and Maddi, Aadyaa and Murakonda, Sasi Kumar and Bindschaedler, Vincent and Shokri, Reza , title =. 2022 , booktitle =

    2022

  56. [56]

    , journal =

    Alexey Kurakin and Natalia Ponomareva, Umar Syed and Liam MacDermed and Andreas Terzis , title=. , journal =

  57. [57]

    Wang and S

    H. Wang and S. Sudalairaj and J. Henning and K. Greenewald and A. Srivastava , booktitle=. Post-processing Private Synthetic Data for Improving Utility on Selected Measures , year=

  58. [58]

    Generating Synthetic Datasets by Interpolating along Generalized Geodesics , year=

    Jiaojiao Fan and David Alvarez-Melis , booktitle=. Generating Synthetic Datasets by Interpolating along Generalized Geodesics , year=

  59. [59]

    2024 , series =

    Donhauser, Konstantin and Abad, Javier and Hulkund, Neha and Yang, Fanny , title =. 2024 , series =

    2024

  60. [60]

    Private prediction for large-scale synthetic text generation

    Amin, Kareem and Bie, Alex and Kong, Weiwei and Kurakin, Alexey and Ponomareva, Natalia and Syed, Umar and Terzis, Andreas and Vassilvitskii, Sergei. Private prediction for large-scale synthetic text generation. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024

    2024

  61. [61]

    , howpublished =

    US Census Bureau , title =. , howpublished =

  62. [62]

    John M. Abowd and Robert Ashmead and Ryan Cumings-Menon and Simson Garfinkel and Micah Heineck and Christine Heiss and Robert Johns and Daniel Kifer and Philip Leclerc and Ashwin Machanavajjhala and Brett Moran and William Sexton and Matthew Spence and Pavel Zhuravlev , title =. , howpublished =

  63. [63]

    CoRR , volume =

    Yinhan Liu and Myle Ott and Naman Goyal and Jingfei Du and Mandar Joshi and Danqi Chen and Omer Levy and Mike Lewis and Luke Zettlemoyer and Veselin Stoyanov , title =. CoRR , volume =. 2019 , url =

    2019

  64. [64]

    Sentence- BERT : Sentence Embeddings using Siamese BERT -Networks

    Reimers, Nils and Gurevych, Iryna. Sentence- BERT : Sentence Embeddings using Siamese BERT -Networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. 2019

    2019

  65. [65]

    Touvron, Hugo and Martin, Louis and Stone, Kevin and Albert, Peter and Almahairi, Amjad and Babaei, Yasmine and Bashlykov, Nikolay and Batra, Soumya and Bhargava, Prajjwal and Bhosale, Shruti and others , journal=

  66. [66]

    Yelp Open Dataset , year =

  67. [67]

    UCT opic: Unsupervised Contrastive Learning for Phrase Representations and Topic Mining

    Li, Jiacheng and Shang, Jingbo and McAuley, Julian. UCT opic: Unsupervised Contrastive Learning for Phrase Representations and Topic Mining. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022

    2022

  68. [68]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Personalized showcases: Generating multi-modal explanations for recommendations , author=. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  69. [69]

    McKenna, Ryan and Mullins, Brett and Sheldon, Daniel and Miklau, Gerome , title =. Proc. VLDB Endow. , month = jul, pages =. 2022 , volume =

    2022

  70. [70]

    International Conference on Machine Learning , pages=

    Generating private synthetic data with genetic algorithms , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  71. [71]

    Advances in Neural Information Processing Systems , volume=

    Iterative methods for private synthetic data: Unifying framework and new methods , author=. Advances in Neural Information Processing Systems , volume=

  72. [72]

    and Lustrek, M

    Vidulin, V. and Lustrek, M. and Kaluza, B. and Pitlaver, R. and Krivec, J. , title =. 2010 , howpublished =

    2010