• Corpus ID: 232240118

A Central Limit Theorem for Differentially Private Query Answering

@inproceedings{Dong2021ACL,
title={A Central Limit Theorem for Differentially Private Query Answering},
author={Jinshuo Dong and Weijie J. Su and Linjun Zhang},
booktitle={Neural Information Processing Systems},
year={2021}
}
• Published in
Neural Information Processing…
15 March 2021
• Computer Science
Perhaps the single most important use case for differential privacy is to privately answer numerical queries, which is usually achieved by adding noise to the answer vector. The central question is, therefore, to understand which noise distribution optimizes the privacy-accuracy trade-off, especially when the dimension of the answer vector is high. Accordingly, an extensive literature has been dedicated to the question and the upper and lower bounds have been successfully matched up to constant…
7 Citations

Figures and Tables from this paper

This work provides a suﬃcient and necessary condition for ( (cid:15), δ )-diﬀerential privacy for all symmetric and log-concave noise densities and demonstrates that this can yield signiﬁcantly lower mean squared errors than those incurred by the currently used Laplace and Gaussian mechanisms.
• Computer Science
ArXiv
• 2022
This paper shows that pure ( cid:15) -DP cannot be decomposed in either way and that there is neither a log-concave CND nor any multivariate CND for (cid: 15) - DP, and shows that Gaussian-DP, (0, δ ) -DP, and Laplace-DP each have both log- Concave and multivariateCNDs.
• Marco Avella-Medina
• Computer Science
Journal of the Royal Statistical Society: Series B (Statistical Methodology)
• 2022
A canonical singleparameter family of privacy notions within the fDP class that is referred to as ‘Gaussian differential privacy’ (GDP), defined based on hypothesis testing of two shifted Gaussian distributions, which is the focal privacy definition among the family of fDP guarantees.
A canonical singleparameter family of privacy notions within the fDP class that is referred to as ‘Gaussian differential privacy’ (GDP), defined based on hypothesis testing of two shifted Gaussian distributions, which is the focal privacy definition among the family of fDP guarantees.
• Computer Science, Mathematics
Journal of the Royal Statistical Society: Series B (Statistical Methodology)
• 2022
The privacy guarantees of any hypothesis testing based definition of privacy (including the original differential privacy definition) converges to GDP in the limit under composition and a Berry–Esseen style version of the central limit theorem is proved, which gives a computationally inexpensive tool for tractably analysing the exact composition of private algorithms.
• Computer Science
ArXiv
• 2021
This rejoinder aims to address two broad issues that cover most comments made in the discussion and discusses some theoretical aspects of the authors' work and comment on how this work might impact the theoretical foundation of privacy-preserving data analysis.

References

SHOWING 1-10 OF 37 REFERENCES

• Computer Science
STOC
• 2014
The results rely on the existence of short fingerprinting codes (Boneh and Shaw, CRYPTO'95; Tardos, STOC'03), which are closely connected to the sample complexity of differentially private data release.
• Mathematics, Computer Science
STOC '12
• 2012
This work gives a mechanism that works unconditionally, and also gives an improved O(log2 d) approximation to the expected l22 error, via a symmetrization argument which argues that there always exists a near optimal differentially private mechanism which adds noise that is independent of the input database.
• Computer Science, Mathematics
STOC
• 2020
New characterizations of the sample complexity of answering linear queries in the local and central models of differential privacy show that a particular factorization mechanism is approximately optimal, and the optimal sample complexity is bounded from above and below by well studied factorization norms of a matrix associated with the queries.
• Computer Science
IEEE Transactions on Information Theory
• 2016
The fundamental tradeoff between privacy and utility in differential privacy is characterized, and the optimal ϵ-differentially private mechanism for a single realvalued query function under a very general utility-maximization (or cost-minimization) framework is derived.
• Computer Science, Mathematics
SIAM J. Comput.
• 2016
This work studies trade-offs between accuracy and privacy in the context of linear queries over histograms and gives an O(\log^2 d) approximation guarantee for the case of \$(\varepsilon,\delta)-approximate differential privacy.
• Computer Science
TCC
• 2006
The study is extended to general functions f, proving that privacy can be preserved by calibrating the standard deviation of the noise according to the sensitivity of the function f, which is the amount that any single argument to f can change its output.
• Computer Science
2014 IEEE International Symposium on Information Theory
• 2014
This work derives the optimal ε-differentially private mechanism for single real-valued query function under a very general utility-maximization (or cost-minimization) framework and concludes that the gains of the staircase mechanism are more pronounced in the moderate-low privacy regime.
• Computer Science, Mathematics
ArXiv
• 2019
It is shown that there is a difference in the privacy loss when the exponential mechanism is chosen adaptively versus non-adaptively, and the best previously known upper bounds for adaptive composition of exponential mechanisms with efficiently computable formulations are improved.
• Computer Science
AISTATS
• 2020
It is shown that the multiplicative gap of the lower bounds and upper bounds goes to zero in various high privacy regimes, proving the tightness of theLower and upper limits and thus establishing the optimality of the truncated Laplacian mechanism.
• Computer Science
2014 IEEE 55th Annual Symposium on Foundations of Computer Science
• 2014
This work provides new algorithms and matching lower bounds for differentially private convex empirical risk minimization assuming only that each data point's contribution to the loss function is Lipschitz and that the domain of optimization is bounded.