pith. sign in

arxiv: 2605.21938 · v1 · pith:MW6D4RKRnew · submitted 2026-05-21 · 💻 cs.LG · cs.CR· cs.IT· math.IT

Optimal Guarantees for Auditing R\'enyi Differentially Private Machine Learning

Pith reviewed 2026-05-22 07:44 UTC · model grok-4.3

classification 💻 cs.LG cs.CRcs.ITmath.IT
keywords Rényi differential privacyblack-box auditingDonsker-Varadhan estimatorminimax lower boundsDP-SGDprivacy leakage estimationhypothesis testing
0
0 comments X

The pith

Auditing Rényi differential privacy with Donsker-Varadhan estimators achieves information-theoretically optimal sample complexity up to logarithmic factors.

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

The paper introduces a hypothesis-testing framework that uses Donsker-Varadhan variational estimators to directly estimate Rényi divergence between neighboring runs of a machine-learning algorithm. This produces explicit non-asymptotic confidence intervals that isolate statistical estimation error from any actual privacy leakage. Matching minimax lower bounds are then shown, establishing that the sample complexity required is nearly optimal in an information-theoretic sense. The result applies to black-box settings and is demonstrated on DP-SGD training of models for MNIST and CIFAR-10 across a range of privacy parameters. If the bounds hold, auditors can now certify claimed Rényi privacy levels with far greater reliability than earlier empirical methods.

Core claim

Class-restricted Donsker-Varadhan estimators, when applied to Rényi divergence estimation via hypothesis testing, deliver non-asymptotic confidence intervals for RDP auditing that separate statistical error from algorithmic leakage; these estimators further satisfy matching minimax lower bounds, proving that the derived sample-complexity guarantees are information-theoretically optimal up to logarithmic factors and thereby supplying the first optimal guarantees for black-box RDP auditing.

What carries the argument

Class-restricted Donsker-Varadhan variational estimator, which recasts Rényi divergence estimation as a variational optimization problem inside a hypothesis test to produce tight, non-asymptotic bounds.

If this is right

  • Auditors obtain rigorous, non-asymptotic guarantees on how tightly a claimed RDP parameter matches the true privacy loss.
  • Empirical lower bounds on privacy leakage improve markedly over prior black-box methods, especially at small and moderate Rényi orders.
  • The same framework applies directly to DP-SGD training pipelines on standard image datasets without requiring internal model access.
  • Sample requirements for reliable auditing scale near-optimally with the desired and the privacy parameter.

Where Pith is reading between the lines

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

  • Regulatory or certification bodies could adopt the method as a standardized test for claimed Rényi privacy levels in deployed models.
  • The variational hypothesis-testing approach may extend to auditing other privacy definitions such as zero-concentrated or concentrated DP.
  • If the black-box separation continues to hold on more complex architectures, the need for white-box or model-specific auditing techniques would diminish.

Load-bearing premise

The Donsker-Varadhan estimator can be realized in a fully black-box manner on any model and data distribution without further restrictions that would couple statistical error to privacy leakage.

What would settle it

An explicit construction or numerical counter-example in which the number of samples needed to certify a given RDP parameter exceeds the paper's upper bound by more than a logarithmic factor, or in which the estimator cannot separate estimation variance from true privacy violation on a simple mechanism such as the Gaussian.

Figures

Figures reproduced from arXiv: 2605.21938 by Benjamin D. Kim, Daniel Alabi, Lav R. Varshney.

Figure 1
Figure 1. Figure 1: Auditing procedure in Algorithm 3. samples, whether existing methods are optimal, and how auditing error should scale with model complexity. Black-box privacy auditing Tight audits for DP-SGD exist in the white-box auditing setting where an adversary can observe and update intermediate gradients [32]. In contrast, in this work we consider a black-box auditing setting, where an adversary can only insert an … view at source ↗
Figure 2
Figure 2. Figure 2: Graph for auditing at α = 1.25 for CIFAR-10 and MNIST datasets on a CNN [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Graph for auditing at α = 2.0 for CIFAR-10 and MNIST datasets on a CNN C Additional Experimental Results This section in the appendix provides additional interpretation of the experimental results reported in Section 4 (in main body) and the accompanying figures. The primary purpose of these experiments is not to propose a new attack on DP-SGD, but rather to validate the theoretical auditing guarantees de￾… view at source ↗
read the original abstract

We study black-box auditing for machine learning algorithms that claim R \ 'enyi differential privacy (RDP) guarantees. We introduce an auditing framework, based on hypothesis testing, that directly estimates R\'enyi divergence between neighboring executions using the Donsker-Varadhan (DV) variational estimator. Our analysis yields explicit and non-asymptotic confidence intervals for RDP auditing via class-restricted DV estimators, separating statistical estimation error from algorithmic privacy leakage. We prove matching minimax lower bounds showing that, up to logarithmic factors, our sample-complexity guarantees are information-theoretically optimal, thereby establishing the first optimal guarantees for auditing RDP via DV estimators. Empirically, we instantiate our framework for auditing DP-SGD in a fully black-box setting. Across MNIST and CIFAR-10, and over a wide range of privacy regimes, our auditors produce a strong overall improvement on empirical RDP lower bounds compared to prior state-of-the-art black-box methods especially at small and moderate R\'enyi orders where accurate auditing is most challenging.

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.

Referee Report

1 major / 3 minor

Summary. The paper introduces a black-box auditing framework for Rényi differentially private ML algorithms based on hypothesis testing and Donsker-Varadhan variational estimators to directly estimate Rényi divergence between neighboring executions. It derives explicit non-asymptotic confidence intervals separating statistical estimation error from algorithmic privacy leakage, proves matching minimax lower bounds (up to logarithmic factors) establishing information-theoretic optimality for the auditing procedure, and reports empirical improvements over prior black-box methods when auditing DP-SGD on MNIST and CIFAR-10 across privacy regimes.

Significance. If the central claims hold, the work delivers the first optimal sample-complexity guarantees for auditing RDP via class-restricted DV estimators. Credit is due for the clean separation of estimation error from leakage, the information-theoretic lower-bound arguments, the explicit non-asymptotic intervals, and the reproducible empirical evaluation on standard datasets that shows gains especially at small and moderate Rényi orders.

major comments (1)
  1. [Theorem 4.3] Theorem 4.3 (minimax lower bound): the lower-bound construction relies on a hypothesis-testing reduction whose test-function class is not shown to coincide with the restricted class F used for the class-restricted DV estimator in the upper bound (Theorem 3.1 and §3.2). Without an explicit argument that the packing or variational lower bound applies to the same F (e.g., bounded neural networks), the claim that the sample-complexity guarantees are information-theoretically optimal for the black-box auditing procedure does not follow.
minor comments (3)
  1. [§3.1] §3.1: the definition of the function class F for the restricted DV estimator is introduced only informally; an explicit statement of the norm or Lipschitz constraint on F in the main text would improve readability.
  2. [Figure 3] Figure 3 caption: the plotted confidence intervals should state the exact Rényi order α and the number of independent runs used to compute the empirical coverage.
  3. [Appendix B.2] Appendix B.2: the covering-number bound used to control the log factor in the sample complexity is stated without a reference to the precise entropy integral or chaining argument employed.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading of the manuscript and for identifying this important point on the alignment between the function classes in the upper and lower bounds. We address the comment below.

read point-by-point responses
  1. Referee: [Theorem 4.3] Theorem 4.3 (minimax lower bound): the lower-bound construction relies on a hypothesis-testing reduction whose test-function class is not shown to coincide with the restricted class F used for the class-restricted DV estimator in the upper bound (Theorem 3.1 and §3.2). Without an explicit argument that the packing or variational lower bound applies to the same F (e.g., bounded neural networks), the claim that the sample-complexity guarantees are information-theoretically optimal for the black-box auditing procedure does not follow.

    Authors: We agree that an explicit connection is required to establish that the minimax lower bound applies to the same restricted class F used by the class-restricted DV estimator. The current manuscript derives the lower bound via a general hypothesis-testing reduction for Rényi divergence estimation. In the revision we will add a short lemma in Section 4 showing that the test functions arising from the packing construction belong to (or can be approximated by elements of) F, under the standard assumption that F consists of bounded neural networks or Lipschitz functions with sufficient capacity. This addition will make the information-theoretic optimality claim precise for the black-box auditing procedure with the restricted class, up to the logarithmic factors already stated. revision: yes

Circularity Check

0 steps flagged

No significant circularity in optimality claims or derivation chain

full rationale

The paper derives its sample-complexity upper bounds and matching minimax lower bounds via separate information-theoretic arguments and variational representations of Rényi divergence. The lower bounds are presented as independent of the specific class-restricted DV estimator instantiations used for the upper bounds and confidence intervals; no equations or steps reduce the claimed optimality to a fitted parameter, self-definition, or self-citation chain by construction. The framework explicitly separates statistical estimation error from algorithmic leakage, and the analysis is self-contained against external benchmarks rather than relying on renaming or smuggling of prior ansatzes.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on standard mathematical properties of Rényi divergence and the Donsker-Varadhan representation; no new free parameters, ad-hoc axioms, or invented entities are introduced in the abstract.

axioms (1)
  • standard math Rényi divergence admits a variational representation via the Donsker-Varadhan formula that can be restricted to suitable function classes.
    Invoked to obtain the class-restricted DV estimator for RDP auditing.

pith-pipeline@v0.9.0 · 5719 in / 1255 out tokens · 53965 ms · 2026-05-22T07:44:00.930022+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    Deep learning with differential privacy

    Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308–318, 2016

  2. [2]

    The existence of error-correcting codes implies privacy lower bounds.IEEE BITS the Information Theory Magazine, pages 1–12, 2026

    Daniel Alabi. The existence of error-correcting codes implies privacy lower bounds.IEEE BITS the Information Theory Magazine, pages 1–12, 2026

  3. [3]

    Privacy and security in distributed data markets

    Daniel Alabi, Sainyam Galhotra, Shagufta Mehnaz, Zeyu Song, and Eugene Wu. Privacy and security in distributed data markets. InCompanion of the 2025 International Conference on Management of Data, SIGMOD/PODS ’25, 2025

  4. [4]

    A variational characterization of rényi divergences.IEEE Transactions on Information Theory, 64(11):6979–6989, 2018

    Venkat Anantharam. A variational characterization of rényi divergences.IEEE Transactions on Information Theory, 64(11):6979–6989, 2018

  5. [5]

    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. InProceedings of the 33rd USENIX Conference on Security Symposium, SEC ’24, USA, 2024. USENIX Association

  6. [6]

    Mutual information neural estimation

    Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajeswar, Sherjil Ozair, Yoshua Bengio, Aaron Courville, and Devon Hjelm. Mutual information neural estimation. InProceedings of the 35th International Conference on Machine Learning (ICML). PMLR, 2018

  7. [7]

    Benjamin Bichsel, Timon Gehr, Dana Drachsler-Cohen, Petar Tsankov, and Martin T. Vechev. Dp-finder: Finding differential privacy violations by sampling and optimization. InProceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security, CCS 2018, Toronto, ON, Canada, October 15-19, 2018, pages 508–524. ACM, 2018

  8. [8]

    Varia- tional representations and neural network estimation of rényi divergences.SIAM Journal on Mathematics of Data Science, 3(4):1093–1116, 2021

    Jeremiah Birrell, Paul Dupuis, Markos A Katsoulakis, Luc Rey-Bellet, and Jie Wang. Varia- tional representations and neural network estimation of rényi divergences.SIAM Journal on Mathematics of Data Science, 3(4):1093–1116, 2021

  9. [9]

    Function-space regularized rényi divergences

    Jeremiah Birrell, Yannis Pantazis, Paul Dupuis, Luc Rey-Bellet, and Markos Katsoulakis. Function-space regularized rényi divergences. InThe Eleventh International Conference on Learning Representations, 2023

  10. [10]

    Cover and Joy A

    Thomas M. Cover and Joy A. Thomas.Elements of Information Theory 2nd Edition (Wiley Series in Telecommunications and Signal Processing). Wiley-Interscience, July 2006

  11. [11]

    Detecting violations of differential privacy

    Zeyu Ding, Yuxin Wang, Guanhong Wang, Danfeng Zhang, and Daniel Kifer. Detecting violations of differential privacy. InProceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security, CCS 2018, Toronto, ON, Canada, October 15-19, 2018, pages 475–489. ACM, 2018

  12. [12]

    Gaussian Differential Privacy

    Jinshuo Dong, Aaron Roth, and Weijie J Su. Gaussian differential privacy.arXiv preprint arXiv:1905.02383, 2019

  13. [13]

    Monroe Donsker and S. R. Srinivasa Varadhan. Asymptotic evaluation of certain markov process expectations for large time, IV.Communications on Pure and Applied Mathematics, 36(2):183–212, 1983

  14. [14]

    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. InEUROCRYPT, 2006

  15. [15]

    Differential privacy in practice: Expose your epsilons!Journal of Privacy and Confidentiality, 9, 10 2019

    Cynthia Dwork, Nitin Kohli, and Deirdre Mulligan. Differential privacy in practice: Expose your epsilons!Journal of Privacy and Confidentiality, 9, 10 2019

  16. [16]

    Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam D. Smith. Calibrating noise to sensitivity in private data analysis. InTCC, 2006. 11

  17. [17]

    Infoshape: Task-based neural data shaping via mutual information

    Homa Esfahanizadeh, William Wu, Manya Ghobadi, Regina Barzilay, and Muriel Médard. Infoshape: Task-based neural data shaping via mutual information. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

  18. [18]

    Detecting violations of differential privacy for quantum algorithms

    Ji Guan, Wang Fang, Mingyu Huang, and Mingsheng Ying. Detecting violations of differential privacy for quantum algorithms. InProceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, CCS 2023, Copenhagen, Denmark, November 26-30, 2023, pages 2277–2291. ACM, 2023

  19. [19]

    Auditing differentially private machine learning: How private is private sgd?Advances in Neural Information Processing Systems, 33:22205–22216, 2020

    Matthew Jagielski, Jonathan Ullman, and Alina Oprea. Auditing differentially private machine learning: How private is private sgd?Advances in Neural Information Processing Systems, 33:22205–22216, 2020

  20. [20]

    The composition theorem for differential privacy

    Peter Kairouz, Sewoong Oh, and Pramod Viswanath. The composition theorem for differential privacy. InInternational conference on machine learning, pages 1376–1385. PMLR, 2015

  21. [21]

    Extremal mechanisms for local differential privacy.Journal of Machine Learning Research, 17(17):1–51, 2016

    Peter Kairouz, Sewoong Oh, and Pramod Viswanath. Extremal mechanisms for local differential privacy.Journal of Machine Learning Research, 17(17):1–51, 2016

  22. [22]

    Cryptanalysis via machine learning based information theoretic metrics.arXiv preprint arXiv:2501.15076, 2025

    Benjamin D Kim, Vipindev Adat Vasudevan, Rafael GL D’Oliveira, Alejandro Cohen, Thomas Stahlbuhk, and Muriel Médard. Cryptanalysis via machine learning based information theoretic metrics.arXiv preprint arXiv:2501.15076, 2025

  23. [23]

    Crypto-mine: Cryptanalysis via mutual information neural estimation

    Benjamin D Kim, Vipindev Adat Vasudevan, Jongchan Woo, Alejandro Cohen, Rafael GL D’Oliveira, Thomas Stahlbuhk, and Muriel Médard. Crypto-mine: Cryptanalysis via mutual information neural estimation. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4820–4824. IEEE, 2024

  24. [24]

    Dp-auditorium: A large-scale library for auditing differential privacy

    William Kong, Andrés Muñoz Medina, Mónica Ribero, and Umar Syed. Dp-auditorium: A large-scale library for auditing differential privacy. In2024 IEEE Symposium on Security and Privacy (SP), pages 110–126, 2024

  25. [25]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

  26. [26]

    The mnist database of handwritten digits.http://yann

    Yann LeCun. The mnist database of handwritten digits.http://yann. lecun. com/exdb/mnist/, 1998

  27. [27]

    MacWilliams and N.J.A

    F.J. MacWilliams and N.J.A. Sloane.The Theory of Error-Correcting Codes. North-holland Publishing Company, 2nd edition, 1978

  28. [28]

    Auditingf-differential privacy in one run,

    Saeed Mahloujifar, Luca Melis, and Kamalika Chaudhuri. Auditing f-differential privacy in one run.arXiv preprint arXiv:2410.22235, 2024

  29. [29]

    Extension of range of functions.Bulletin of the American Mathemati- cal Society, 40:837–842, 1934

    Edward James McShane. Extension of range of functions.Bulletin of the American Mathemati- cal Society, 40:837–842, 1934

  30. [30]

    Rényi differential privacy

    Ilya Mironov. Rényi differential privacy. In2017 IEEE 30th Computer Security Foundations Symposium (CSF). IEEE, 2017

  31. [31]

    Nearly tight black-box auditing of differentially private machine learning.Advances in Neural Information Processing Systems, 37:131482–131502, 2024

    Meenatchi Sundaram Muthu Selva Annamalai and Emiliano De Cristofaro. Nearly tight black-box auditing of differentially private machine learning.Advances in Neural Information Processing Systems, 37:131482–131502, 2024

  32. [32]

    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. In32nd USENIX Security Symposium (USENIX Security 23), pages 1631–1648, 2023

  33. [33]

    Membership inference attacks against machine learning models

    Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In2017 IEEE Symposium on Security and Privacy, SP 2017, San Jose, CA, USA, May 22-26, 2017, pages 3–18. IEEE Computer Society, 2017

  34. [34]

    Privacy auditing with one (1) training run

    Thomas Steinke, Milad Nasr, and Matthew Jagielski. Privacy auditing with one (1) training run. InAdvances in Neural Information Processing Systems (NeurIPS), volume 36, 2023. 12

  35. [35]

    Analytic extensions of differentiable functions defined in closed sets.Trans- actions of the American Mathematical Society, 36(1):63–89, 1934

    Hassler Whitney. Analytic extensions of differentiable functions defined in closed sets.Trans- actions of the American Mathematical Society, 36(1):63–89, 1934

  36. [36]

    Enhanced membership inference attacks against machine learning models

    Jiayuan Ye, Aadyaa Maddi, Sasi Kumar Murakonda, Vincent Bindschaedler, and Reza Shokri. Enhanced membership inference attacks against machine learning models. In Heng Yin, Angelos Stavrou, Cas Cremers, and Elaine Shi, editors,Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, CCS 2022, Los Angeles, CA, USA, November 7-1...

  37. [37]

    These results justify when empirical audits can certify both privacy violations and near-tight compliance

    How bounded privacy loss enables two-sided confidence intervals.By imposing a bounded privacy-loss assumption (exact for pure DP and a reasonable high-probability surrogate for many RDP mechanisms) we derive finite-sample upper and lower confidence bounds using classical concentration inequalities. These results justify when empirical audits can certify b...

  38. [38]

    This formally validates the use of neural divergence estimators as statistically principled auditing tools, rather than heuristic approximations

    Why the DV-based estimator is statistically sound for auditing.We provide a full non- asymptotic analysis of the DV variational estimator over restricted critic classes, showing uniform convergence, explicit rates, and consistency. This formally validates the use of neural divergence estimators as statistically principled auditing tools, rather than heuri...

  39. [39]

    (Uniform boundedness)sup θ,z |Tθ(z)| ≤M

  40. [40]

    (Lipschitz parameterization) For allz∈Ω, |Tθ(z)−T θ′(z)| ≤L∥θ−θ ′∥. Define the population DV Rényi functional V(θ) := 1 α−1 logE Q h e(α−1)Tθ(X) i − 1 α logE P h eαTθ(Y) i , and its empirical estimator bVn(θ) := 1 α−1 log 1 n nX i=1 e(α−1)Tθ(Xi) ! − 1 α log 1 n nX i=1 eαTθ(Yi) ! . Define RΘ α (Q∥P) := sup θ∈Θ V(θ), bRΘ α,n(Q∥P) := sup θ∈Θ bVn(θ). Then for...