pith. sign in

arxiv: 2605.05723 · v1 · submitted 2026-05-07 · 💻 cs.CR

α-Wasserstein Mechanism for R\'{e}nyi Pufferfish Privacy

Ni Ding , Wenjin Yang , Zijian Zhang This is my paper

Pith reviewed 2026-05-08 09:33 UTC · model grok-4.3

classification 💻 cs.CR
keywords Rényi Pufferfish PrivacyWasserstein mechanismLaplace mechanismGaussian mechanismHölder's inequalitydifferential privacyprivacy mechanisms
0
0 comments X

The pith

An upper bound on the α-Wasserstein metric calibrates Laplace and Gaussian noise scales to achieve exact (α, ε)-Rényi Pufferfish Privacy.

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

This paper develops the α-Wasserstein mechanism to protect private information under the Rényi Pufferfish Privacy definition by adding calibrated Laplace or Gaussian noise. The authors apply Hölder's inequality to derive an upper bound on the α-Wasserstein distance between distributions and use it to set the noise scale parameter so that the released output meets the privacy condition for any α greater than 1. The same bound recovers the known W_∞ mechanism for standard Pufferfish privacy in the limit as α goes to infinity and extends to the exponential mechanism plus a Gaussian version that generalizes Rényi differential privacy. Experiments indicate the resulting mechanisms require less noise power than the conventional infinite-α approach, with the Gaussian variant delivering higher utility. A reader would care because Pufferfish privacy lets users specify exactly which facts need protection rather than protecting all data, and tighter noise calibration preserves more accuracy in statistical releases from sensitive sources.

Core claim

By leveraging Hölder's inequality, the scale parameter of the Laplace mechanism can be calibrated via an upper bound on the W_α metric to satisfy (α, ε)-Rényi Pufferfish Privacy for α ∈ (1, ∞]. At the limit α = ∞ this framework recovers the established W_∞ mechanism for ε-pufferfish privacy. The result is extended to the exponential mechanism. A W_α mechanism is also proposed for Gaussian noise for α ∈ (1, ∞), demonstrating that it generalizes existing results within the Rényi Differential Privacy framework. The mechanisms achieve exact (α, ε)-Rényi Pufferfish Privacy without requiring additional relaxations such as δ-approximations.

What carries the argument

The α-Wasserstein mechanism, which sets the scale of added Laplace or Gaussian noise using an upper bound on the W_α metric obtained from Hölder's inequality.

If this is right

  • The α-Wasserstein mechanism achieves exact (α, ε)-Rényi Pufferfish Privacy without δ-approximations.
  • The framework recovers the established W_∞ mechanism for ε-pufferfish privacy as α approaches infinity.
  • The Gaussian W_α mechanism generalizes existing results in the Rényi Differential Privacy framework.
  • Experimental evaluations show significantly reduced noise power compared to the W_∞-based approach, with the Gaussian mechanism providing superior utility over the Laplace mechanism.

Where Pith is reading between the lines

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

  • The calibration technique could be tested on other noise families or privacy definitions that rely on similar metric bounds between output distributions.
  • In practice the noise reduction might be measured on datasets with restricted secret spaces to quantify real-world utility gains over existing methods.
  • Links to optimal transport could yield tighter bounds or entirely new mechanisms for privacy-preserving data release.

Load-bearing premise

An upper bound on the W_α metric derived via Hölder's inequality is sufficient to guarantee the exact (α, ε)-Rényi Pufferfish Privacy definition without hidden gaps or extra assumptions on the secret space or data distributions.

What would settle it

A concrete counterexample distribution and secret pair where the noise scale chosen from the W_α upper bound still fails to satisfy the (α, ε)-Rényi Pufferfish Privacy inequality.

Figures

Figures reproduced from arXiv: 2605.05723 by Ni Ding, Wenjin Yang, Zijian Zhang.

Figure 1
Figure 1. Figure 1: Experimental results using adult, heart disease and student performance datasets from UCI machine learning repository [29]: rows 1-3 compare Theorems 1 and 2 to W∞ based mechanism in [16, Corollary 3.1] for ϵ = 0.5, 1; rows 4-5 show noise reduction by Gaussian mechanism in Theorem 2 as compared to Laplace mechanism in Theorems 1. Range α ∈ (0, 1): It is not difficult to derive the sufficient condition R e … view at source ↗
Figure 2
Figure 2. Figure 2: The comparison of Wα(PX|si,ρ, PX|sj ,ρ) with it’s upper bound e α−1 α ϵ in (14) for fixed ϵ. 13 view at source ↗
Figure 3
Figure 3. Figure 3: The variation of scale parameter b determined by Theorem 1 and [16, Corollary 3.1]. They both approach W∞/ϵ mechanism proposed in [9] as α → ∞ for attaining ϵ-pufferfish privacy view at source ↗
read the original abstract

This paper introduces the $\alpha$-Wasserstein mechanism for achieving R\'{e}nyi Pufferfish Privacy using Laplace and Gaussian noise. By leveraging H\"{o}lder's inequality, we demonstrate that the scale parameter of the Laplace mechanism can be calibrated via an upper bound on the $W_\alpha$ metric to satisfy $(\alpha, \epsilon)$-R\'{e}nyi Pufferfish Privacy for $\alpha \in (1, \infty]$. We show that at the limit $\alpha = \infty$, this framework recovers the established $W_\infty$ mechanism for $\epsilon$-pufferfish privacy. This result is subsequently extended to the exponential mechanism. Furthermore, we propose a $W_\alpha$ mechanism for Gaussian noise for $\alpha \in (1, \infty)$, demonstrating that it generalizes existing results within the R\'enyi Differential Privacy framework. Experimental evaluations reveal that our $\alpha$-Wasserstein mechanism significantly reduces noise power compared to the conventional $W_\infty$-based approach, with the Gaussian mechanism providing superior utility over the Laplace mechanism. Notably, the mechanisms derived in this work achieve exact $(\alpha, \epsilon)$-R\'{e}nyi Pufferfish Privacy without requiring additional relaxations, such as $\delta$-approximations.

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

2 major / 2 minor

Summary. The manuscript introduces the α-Wasserstein mechanism for (α, ε)-Rényi Pufferfish Privacy. It applies Hölder's inequality to derive an upper bound on the α-Wasserstein distance W_α between secret-pair distributions, then uses this bound to calibrate the scale parameter of the Laplace mechanism (for α ∈ (1, ∞]) and a Gaussian mechanism (for α ∈ (1, ∞)) so that the resulting noise addition satisfies the Rényi Pufferfish definition. The framework recovers the known W_∞ mechanism for ε-Pufferfish privacy in the α → ∞ limit, extends the approach to the exponential mechanism, and generalizes existing Rényi differential privacy results for the Gaussian case. Experiments on synthetic data are reported to show that the proposed mechanisms require substantially less noise power than the conventional W_∞ baseline, with the Gaussian variant outperforming Laplace.

Significance. If the central calibration argument is rigorous and closes without hidden gaps, the work supplies a parameterized family of mechanisms that can exploit finite α-Wasserstein distances rather than worst-case W_∞ distances, offering a principled route to improved utility while retaining an exact (α, ε) guarantee without δ-relaxation. The recovery of the established W_∞ result and the generalization of RDP-style Gaussian mechanisms are technically attractive features.

major comments (2)
  1. [Abstract / main theorem] Abstract and main calibration argument: The claim that an upper bound on W_α obtained via Hölder's inequality is sufficient to set the Laplace/Gaussian scale so that D_α(M(D_s) || M(D_{s'})) ≤ ε holds exactly for arbitrary secret-pair distributions is load-bearing. Hölder's inequality supplies a sufficient but possibly strict bound; if the inequality is not tight for the optimal coupling or if the secret space violates implicit integrability/compactness conditions, the resulting scale may fail to enforce the stated Rényi divergence bound. The manuscript must explicitly derive the composition of the W_α bound into the Rényi divergence (including any assumptions on the prior or metric space) rather than asserting exactness from the upper bound alone.
  2. [Gaussian mechanism section] § on Gaussian mechanism: The extension to Gaussian noise for α ∈ (1, ∞) is stated to generalize existing RDP results, yet the precise relationship between the W_α-calibrated scale and the standard RDP Gaussian scale (which is typically derived from the 2-Wasserstein or variance) is not shown to be tight or to preserve the exact (α, ε) guarantee under the same Hölder-derived bound. A concrete comparison of the two scales and the resulting privacy parameters is required.
minor comments (2)
  1. [Experiments] The experimental section should report the concrete secret-pair distributions, how W_α is estimated or bounded in practice, and statistical variability across runs; the current description of 'significantly reduces noise power' is too qualitative to assess reproducibility.
  2. [Preliminaries] Notation for the α-Wasserstein distance and the resulting scale parameter should be introduced with an explicit equation early in the paper to avoid ambiguity when the same symbol is later used for the calibrated noise variance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the presentation of our calibration arguments. We address each major point below and will revise the manuscript accordingly to make the derivations explicit and add the requested comparisons.

read point-by-point responses

  1. Referee: Abstract / main theorem: The claim that an upper bound on W_α obtained via Hölder's inequality is sufficient to set the Laplace/Gaussian scale so that D_α(M(D_s) || M(D_{s'})) ≤ ε holds exactly for arbitrary secret-pair distributions is load-bearing. Hölder's inequality supplies a sufficient but possibly strict bound; if the inequality is not tight for the optimal coupling or if the secret space violates implicit integrability/compactness conditions, the resulting scale may fail to enforce the stated Rényi divergence bound. The manuscript must explicitly derive the composition of the W_α bound into the Rényi divergence (including any assumptions on the prior or metric space) rather than asserting exactness from the upper bound alone.

    Authors: We agree that the derivation steps from the Hölder bound on W_α to the Rényi divergence bound D_α ≤ ε should be spelled out explicitly rather than left implicit. In the revised manuscript we will insert a dedicated lemma (or expanded proof paragraph) that starts from the definition of the α-Wasserstein distance, applies the Hölder-derived upper bound, and then invokes the known relationship between Wasserstein distance and Rényi divergence for the Laplace (or Gaussian) mechanism under the Pufferfish secret-pair setting. We will state the standing assumptions: the underlying metric space is Polish, the secret-pair distributions have finite α-moments, and the prior is a probability measure on the secret space. Because the bound is sufficient (not necessarily tight), the calibrated scale guarantees D_α ≤ ε, possibly with conservative noise; we will clarify this distinction and note that the guarantee remains exact (no δ) under these conditions. revision: yes

  2. Referee: Gaussian mechanism section: The extension to Gaussian noise for α ∈ (1, ∞) is stated to generalize existing RDP results, yet the precise relationship between the W_α-calibrated scale and the standard RDP Gaussian scale (which is typically derived from the 2-Wasserstein or variance) is not shown to be tight or to preserve the exact (α, ε) guarantee under the same Hölder-derived bound. A concrete comparison of the two scales and the resulting privacy parameters is required.

    Authors: We will add an explicit comparison in the revised Gaussian-mechanism section. For the special case α = 2 our W_α-calibrated scale reduces (up to a universal constant) to the standard RDP Gaussian scale obtained from the 2-Wasserstein distance; we will derive the exact algebraic relation between the two expressions. For general α we will show that the same Hölder bound yields a scale that satisfies the (α, ε)-Rényi Pufferfish guarantee exactly, and we will include a short table contrasting the resulting noise variances for representative α values against the classical RDP formula. This addition will also highlight that the Pufferfish formulation recovers the RDP result when the secret space collapses to neighboring datasets. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation uses external Hölder's inequality for sufficient bound on W_α

full rationale

The paper's central step calibrates Laplace/Gaussian scale from an upper bound on W_α obtained via Hölder's inequality to ensure the Rényi divergence between mechanism outputs is at most ε for secret pairs. This is a one-way sufficient condition derived from standard analysis, not a self-definition where the privacy parameter is fitted from or defined in terms of the same quantity. The limit case α→∞ recovers the known W_∞ mechanism without re-deriving it from the paper's own inputs. No fitted parameters are relabeled as predictions, no load-bearing self-citations close the argument, and the mechanism is shown to achieve exact (α,ε)-Rényi Pufferfish Privacy without δ-relaxations. The derivation chain remains independent of its target result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on Hölder's inequality (standard math) to obtain the noise scale bound and on the existing definitions of Rényi Pufferfish Privacy and Wasserstein metrics; no new free parameters, ad-hoc axioms, or invented entities are introduced in the abstract.

axioms (1)
  • standard math Hölder's inequality
    Invoked to produce an upper bound on the W_α metric that calibrates the Laplace scale parameter.

pith-pipeline@v0.9.0 · 5535 in / 1364 out tokens · 82270 ms · 2026-05-08T09:33:37.355388+00:00 · methodology

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

41 extracted references · 41 canonical work pages · 1 internal anchor

  1. [1]

    McSherry, K

    Dwork, C., F. McSherry, K. Nissim, et al. Calibrating noise to sensitivity in private data analysis. In S. Halevi, T. Rabin, eds.,Theory of Cryptography, pages 265–284. Springer Berlin Heidelberg, Berlin, Heidelberg, 2006

    work page 2006

  2. [2]

    Differential privacy

    Dwork, C. Differential privacy. In M. Bugliesi, B. Preneel, V . Sassone, I. Wegener, eds.,Automata, Languages and Programming, pages 1–12. Springer Berlin Heidelberg, Berlin, Heidelberg, 2006

    work page 2006

  3. [3]

    Wasserman, L., S. Zhou. A statistical framework for differential privacy.Journal of the American Statistical Association, 105(489):375–389, 2010

    work page 2010

  4. [4]

    Abowd, J. M. The u.s. census bureau adopts differential privacy. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’18, pages 2867–2867. ACM, 2018

    work page 2018

  5. [5]

    Abadi, M., A. Chu, I. Goodfellow, et al. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, pages 308–318. ACM, 2016

    work page 2016

  6. [6]

    Vejdanihemmat, M

    Mohammadi, M., M. Vejdanihemmat, M. Lotfinia, et al. Differential privacy for deep learning in medicine. arXiv e-prints, pages arXiv–2506, 2025

    work page 2025

  7. [7]

    Machanavajjhala

    Kifer, D., A. Machanavajjhala. A rigorous and customizable framework for privacy. InProceedings of the 31st ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems, PODS ’12, page 77–88. Association for Computing Machinery, New York, NY , USA, 2012

    work page 2012

  8. [8]

    Pufferfish: A framework for mathematical privacy definitions.ACM Transactions on Database Systems, 39(1), 2014

    —. Pufferfish: A framework for mathematical privacy definitions.ACM Transactions on Database Systems, 39(1), 2014

    work page 2014

  9. [9]

    Song, S., Y . Wang, K. Chaudhuri. Pufferfish privacy mechanisms for correlated data. InProceedings of the 2017 ACM International Conference on Management of Data, page 1291–1306. New York, NY , USA, 2017

    work page 2017

  10. [10]

    De Pascale, P

    Champion, T., L. De Pascale, P. Juutinen. The ∞-Wasserstein distance: Local solutions and existence of optimal transport maps.SIAM Journal on Mathematical Analysis, 40(1):1–20, 2008

    work page 2008

  11. [11]

    De Pascale, L., J. Louet. A study of the dual problem of the one-dimensional l∞-optimal transport problem with applications.Journal of Functional Analysis, 276(11):3304–3324, 2019

    work page 2019

  12. [12]

    Kantorovich mechanism for pufferfish privacy

    Ding, N. Kantorovich mechanism for pufferfish privacy. In G. Camps-Valls, F. J. R. Ruiz, I. Valera, eds., Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, vol. 151 of Proceedings of Machine Learning Research, pages 5084–5103. PMLR, 2022

    work page 2022

  13. [13]

    Domingo-Ferrer, D

    Soria-Comas, J., J. Domingo-Ferrer, D. Sanchez, et al. Individual differential privacy: A utility-preserving formulation of differential privacy guarantees.IEEE Transactions on Information Forensics and Security, 12(6):1418–1429, 2017

    work page 2017

  14. [14]

    Li, B., W. Wang, P. Ye. The limits of differential privacy in online learning. InAdvances in Neural Information Processing Systems 37, NeurIPS 2024, pages 65328–65360. Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2024

    work page 2024

  15. [15]

    Rényi differential privacy

    Mironov, I. Rényi differential privacy. In2017 IEEE 30th Computer Security Foundations Symposium (CSF), pages 263–275. 2017

    work page 2017

  16. [16]

    Bellet, M

    Pierquin, C., A. Bellet, M. Tommasi, et al. Rényi Pufferfish Privacy: General Additive Noise Mechanisms and Privacy Amplification by Iteration via Shift Reduction Lemmas. InInternational Conference on Machine Learning (ICML 2024). Vienna (Austria), Austria, 2024

    work page 2024

  17. [17]

    On measures of entropy and information

    Rényi, A. On measures of entropy and information. InProceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics, vol. 4, pages 547–562. University of California Press, 1961

    work page 1961

  18. [18]

    Harremoes

    van Erven, T., P. Harremoes. Rényi divergence and Kullback-Leibler divergence.IEEE Transactions on Information Theory, 60(7):3797–3820, 2014

    work page 2014

  19. [19]

    Villani, C.Optimal transport: old and new, vol. 338. Springer, 2009

    work page 2009

  20. [20]

    Optimal transport for applied mathematicians.Birkäuser, NY, 55(58-63):94, 2015

    Santambrogio, F. Optimal transport for applied mathematicians.Birkäuser, NY, 55(58-63):94, 2015. 10

    work page 2015

  21. [21]

    Yang, W., N. Ding, Z. Zhang, et al. Noise reduction for pufferfish privacy: A practical noise calibration method.arXiv preprint arXiv:2601.06385, 2026

    work page arXiv 2026

  22. [22]

    Ding, N., S. Lu, W. Yang, et al. Multi-user pufferfish privacy.arXiv preprint arXiv:2512.18632, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Brent, R. P. An algorithm with guaranteed convergence for finding a zero of a function.The Computer Journal, 14(4):422–425, 1971

    work page 1971

  24. [24]

    Süli, E., D. F. Mayers.An introduction to numerical analysis. Cambridge university press, 2003

    work page 2003

  25. [25]

    Mironov, K

    Feldman, V ., I. Mironov, K. Talwar, et al. Privacy amplification by iteration. In2018 IEEE 59th Annual Symposium on Foundations of Computer Science (FOCS), pages 521–532. IEEE, 2018

    work page 2018

  26. [26]

    Roth, et al

    Dwork, C., A. Roth, et al. The algorithmic foundations of differential privacy.Found. Trends Theor. Comput. Sci., 9(3-4):211–407, 2014

    work page 2014

  27. [27]

    Balle, B., Y .-X. Wang. Improving the Gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. In J. Dy, A. Krause, eds.,Proceedings of the 35th International Conference on Machine Learning, vol. 80 ofProceedings of Machine Learning Research, pages 394–403. PMLR, 2018

    work page 2018

  28. [28]

    arXiv preprint arXiv:1908.10530 (2019)

    Mironov, I., K. Talwar, L. Zhang. Rényi differential privacy of the sampled gaussian mechanism.arXiv preprint arXiv:1908.10530, 2019

    work page arXiv 1908

  29. [29]

    Asuncion, A., D. Newman. UCI machine learning repository https://archive.ics.uci.edu/ml/index.php, 2007

    work page 2007

  30. [30]

    Esposito, A. R., M. Gastpar, I. Issa. Robust generalization via f-mutual information. pages 2723–2728, 2020

    work page 2020

  31. [31]

    Generalization error bounds via Rényi-, f-divergences and maximal leakage.IEEE Transactions on Information Theory, 67(8):4986–5004, 2021

    —. Generalization error bounds via Rényi-, f-divergences and maximal leakage.IEEE Transactions on Information Theory, 67(8):4986–5004, 2021

    work page 2021

  32. [32]

    Information theoretic proofs of entropy power inequalities.IEEE Transactions on Information Theory, 57(1):33–55, 2011

    Rioul, O. Information theoretic proofs of entropy power inequalities.IEEE Transactions on Information Theory, 57(1):33–55, 2011

    work page 2011

  33. [33]

    Rényi entropy power and normal transport

    —. Rényi entropy power and normal transport. In2020 International Symposium on Information Theory and Its Applications (ISITA), pages 1–5. 2020

    work page 2020

  34. [34]

    Guessing and entropy

    Massey, J. Guessing and entropy. InProceedings of 1994 IEEE International Symposium on Information Theory, ISIT-94, page 204. IEEE

    work page 1994

  35. [35]

    An inequality on guessing and its application to sequential decoding.IEEE Transactions on Information Theory, 42(1):99–105, 1996

    Arikan, E. An inequality on guessing and its application to sequential decoding.IEEE Transactions on Information Theory, 42(1):99–105, 1996

    work page 1996

  36. [36]

    Kosut, L

    Liao, J., O. Kosut, L. Sankar, et al. Tunable measures for information leakage and applications to privacy-utility tradeoffs.IEEE Transactions on Information Theory, 65(12):8043–8066, 2019

    work page 2019

  37. [37]

    Farokhi, T

    Ding, N., F. Farokhi, T. Guo, et al.α-leakage interpretation of sibson mutual information and rényi capacity. In2025 IEEE Information Theory Workshop (ITW), pages 752–757. IEEE, 2025

    work page 2025

  38. [38]

    Dowson, D., B. Landau. The fréchet distance between multivariate normal distributions.Journal of multivariate analysis, 12(3):450–455, 1982

    work page 1982

  39. [39]

    Givens, C. R., R. M. Shortt. A class of Wasserstein metrics for probability distributions.Michigan Mathematical Journal, 31(2):231–240, 1984

    work page 1984

  40. [40]

    Wasserstein geometry of Gaussian measures.Osaka Journal of Mathematics, 48(4):1005–1026, 2011

    Takatsu, A. Wasserstein geometry of Gaussian measures.Osaka Journal of Mathematics, 48(4):1005–1026, 2011

    work page 2011

  41. [41]

    Daoxiang, Z., P. Yan. On the hardy–carleman inequality for a negative exponent.Journal of mathematical inequalities, 11(3):885–890, 2017. 11 A Proof of Corollary 1 Proof. The proof is similar to Theorem 1. We still apply the Hölder’s inequality, but use the triangular inequality,P Nθ (y−x)≤e η(θ)c(x−x ′)PNθ (y−x ′),∀x, x ′, y. Dα(PY|s i,ρ∥PY|s j ,ρ) = 1 α...

    work page 2017