Adaptive Privacy Composition for Accuracy-first Mechanisms

In many practical applications of differential privacy, practitioners seek to provide the best privacy guarantees subject to a target level of accuracy. A recent line of work by Ligett et al. '17 and Whitehouse et al. '22 has developed such accuracy-first mechanisms by leveraging the idea of noise reduction that adds correlated noise to the sufficient statistic in a private computation and produces a sequence of increasingly accurate answers. A major advantage of noise reduction mechanisms is that the analysts only pay the privacy cost of the least noisy or most accurate answer released. Despite this appealing property in isolation, there has not been a systematic study on how to use them in conjunction with other differentially private mechanisms. A fundamental challenge is that the privacy guarantee for noise reduction mechanisms is (necessarily) formulated as ex-post privacy that bounds the privacy loss as a function of the released outcome. Furthermore, there has yet to be any study on how ex-post private mechanisms compose, which allows us to track the accumulated privacy over several mechanisms. We develop privacy filters [Rogers et al. '16, Feldman and Zrnic '21, and Whitehouse et al. '22'] that allow an analyst to adaptively switch between differentially private and ex-post private mechanisms subject to an overall differential privacy guarantee.

Paper

References (23)

Scroll for more · 11 remaining

Similar papers

Peer review

Reviewer dJHg6/10 · confidence 3/52023-07-03

Summary

This paper presents extensive background on ex-post-privacy as well as some new analysis providing a new method for leveraging this framework for the design of differentially private interactive protocols. Finally, this framework is instantiated in a method for releasing private counts while targeting an accuracy constraint.

Strengths

* The theoretical framework presented is quite general and seems to be very strong. * Though relaxed notions of privacy are used internally in mechanism design, the main results here are stated for more tradition $(\epsilon, \delta)$ DP. * The power of the theoretical framework _could_ conceviably lead to quite substantial concrete implementations and applications.

Weaknesses

* The generality of the analysis yields an associated weakness: for the vast majority of the main body, the precise methods of application are somewhat unclear. E.g. it is not obvious until the end of the paper that we will not literally be guaranteeing that all of our released values are withing percent, which makes the design of interactive protocols which use this style of mechanism + composition an important and more or less unstated problem space. * The point above can be restated in more direct terms: if the authors wish to see this analysis and framework have a big impact, it is probably still to much of a lift to ask the community to fill in the gaps between the abstract presentation which dominates here and concrete algorithmic implementations.

Questions

* IIUC, the final guarantee claimed here is epsilon, delta DP. Is this correct? If so, I think the paper could benefit from a stronger statement of this much earlier. E.g. the abstract uses ‘differentially private’ and ‘ex post private’, but the statement about an overall guarantee just uses the word ‘privacy’. This leaves a reader somewhat confused about which notion of privacy applies here. * In theorem 3, what _exactly_ do we mean by statistic and its sensitivity? IE, a Brownian motion itself has no data-dependence, but a Brownian noise reduction mechanism does. Is the notion of statistic and sensitivity we are using here something like: the Brownian noise-reduction mechanism is data-dependent, so its path is data-dependent, and the statistic is a function of the path--so for every neighboring $x, x'$, the function $f_i$ of these different paths can differ by no more than 1, _uniformly_ over the randomness in the BM? This would seem to me to be a somewhat different notion than that of usual sensitivity, where there is no randomness involved. Perhaps the term 'statistic' is intended to imply that the randomness has already been integrated out? But if so, what does this mean precisely (e.g. what would an example be)? Further, it seems like the result of Theorem 3 is independent of these statistics $f_i$, unless they are hiding in the stopping times $T_i$ somewhere? NOTE: it is possible that this $f_i$ is intended to recall $f$ of definition 4.4; but if so, I am not quite grokking this relationship--since the $f$ in definition 4.4 can, I guess, be viewed as one _particular_ statistic of the function $BM(x)$ (the limit of the expectation of the second toordinate as $t \to 0$), but doesn't really provide any immediate pointers to generalizing this notion? * It's not necessarily clear from the presentation of the paper what relationship there is between ex-post mechanisms, noise-reduction mechanisms, and the Brownian noise reduction mechanism. Meaning, e.g. the statement in definition 4.2 seems to apply to the privacy loss random variable directly, considered as a function of both its inputs and its randomness--at least, this is how I interpret the equality in definition 2, equality of functions. If this is true, seemingly this definition would be applicable to any particular property of the functions $L$; e.g. if $\cal{L}^{(k)}$ was $(\epsilon, \delta)$-DP then $\cal{L}^{(1:k)}$ would be as well (since it is just a statement about the function $L$, and these functions are equal). But thinking about the domains and ranges, it seems like these functions take values in different spaces. So it's not clear, I suppose, what this equality means? Maybe there is a suppressed relationship to ex-post privacy? * If I read correctly the definition of $(\cal{E}, \delta, \epsilon^*)$ ex-post-private (internal to definition 2) is equivalent to $(\epsilon^*, \delta)$ DP. Is this correct? If so, readers may find it useful to have this explicitly called out, since likely many are more familiar with DP than ex-post privacy. * Nit on eqn (4): IIUC, choosing the particular functional form here (e.g. $\frac{1}{2 t_n^{(kn)}}$) doesn't necessarily play an enormous role--we could likely choose some 'other way' of determining how much 'budget is left' and push the same choice into Lemma 5.2 (and Theorem 4). Is this correct, or is there something fundamental about the way this function is expressed? If there is nothing fundamental, I would advocate as a reader to abstract over this function, and clarify what properties are 'actually' needed here. * Nit on conclusion of Thm 4. Can this be stated as a zCDP guarantee? Or are we doing something _specific_ that forces us to go to $(\epsilon, \delta)$-DP?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The major unaddressed limitation is the potential difficulty of designing effective algorithms which leverage this perspective and analysis. That is, it is not entirely clear how fruitful the analysis presented here will be. This paper _could_ be _extremely_ strong with some reasonable extensions, perhaps as a theory paper with associated 'mechanism design paper' that provides many instantiations of this theoretical framework to perform various privacy sensitive tasks. For this reason, the paper reads as somewhat borderline to me--potentially pointing the way to serious applications, but not quite closing the gap entirely yet. Negative social impacts not immediately applicable.

Reviewer njjx5/10 · confidence 3/52023-07-05

Summary

This paper tries to tackle the following basic question in differential privacy: suppose I desire to answer as many as possible counting queries with low **relative error** with a predetermined privacy budget (\eps,\delta). What is the best mechanism I can use? Building on prior works on "ex-post" private mechanisms (Whitehouse et al '22), this paper makes the following contributions: (1) An improved analysis for ex-post privacy mechanisms (in particular the Brownian Noise Reduction Mechanism); (2) A "Privacy filter" which allows the user to compose concentrated-DP mechanisms and the Brownian mechanism adaptively. (3) An experimental section, demonstrating the improvement of the proposed algorithm, compared with the baseline approach of "doubling". It appears that, compared with the baseline approach, the new algorithm can answer twice as many queries with the same privacy budget and accuracy requirement.

Strengths

* The new analysis for the Brownian mechanism looks nice. It can potentially make the technique more useful. * The experiment setup is thoroughly described, and the improvement looks significant.

Weaknesses

* It looks like the composition for ex-post mechanisms is omitted in the main body. However, you claimed this as one of your main contributions. * Why only analyze the composition of concentrated-DP mechanisms with Brownian mechanisms? This seems to make the theory less general. * From the practice side, this paper only runs experiments on synthetic data sets, which might not be convincing enough to a broader community.

Questions

* You conjectured the composition for ex-post mechanisms is linear in general. I was wondering why it is the case. * How does this ex-post private approach compare with the (private) hyper-parameter selection approach? See, for example, here: https://arxiv.org/abs/2301.00301

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

NA

Reviewer hK6N5/10 · confidence 2/52023-07-07

Summary

‘Accurary-first’ mechanisms try to achieve the strongest possible privacy guarantee that satisfies a target level of accuracy. One such example is a noise reduction mechanism, which scales up the privacy cost / decreases the noise level until the target utility is achieved. Noise reduction mechanisms only pay the privacy cost of the final noise addition, but provide ex-post privacy guarantees whose composition bounds are unknown and which are not compatible with DP. This paper provides composition bounds for ex-post DP, and develops a unified privacy filter that can switch between DP and ex-post DP mechanisms subject to an overall privacy guarantee.

Strengths

1. Very interesting and strong potential for impact. Seems like it could be a big step towards making DP more practical! 2. Novel privacy filter and analysis of the ex-post privacy guarantee for a Brownian filter.

Weaknesses

The paper was very dense and difficult to read, which I feel limits its appeal. (Brownian motion / the Brownian mechanism, for example, are never really explained.) I think the paper would benefit greatly from being gentler to its readers.

Questions

The paper focuses very much on composition bounds for ex-post privacy, but I think that there should also be a discussion on post-processing. The way that Definition 4.1 isn’t defined on sets, it seems like it wouldn’t be closed under post-processing. So I think that converting this into $(\epsilon, \delta)$-DP is an important technical detail that should be discussed in the paper. On a similar note — why the overloading of notation in Definition 4.1? It was unclear to me which version of ‘ex-post’ was being used throughout the remainder of the paper. And I wasn’t sure of what was the purpose of imposing a bound $\epsilon^*$ on the ex-post privacy loss. Isn’t that just redefining DP but without sets?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

1 poor

Contribution

4 excellent

Limitations

No issues.

Reviewer HnZw7/10 · confidence 2/52023-07-07

Summary

This paper focuses on the study of adaptive compositions of ex-post differential privacy, where privacy guarantees depend on the outcome of the algorithm. The utilization of ex-post DP enables analysts to achieve a more favorable privacy-utility trade-off. The paper introduces a tool for composing ex-post DP and traditional DP, providing a unified privacy filter that combines the Brownian Mechanism with traditional DP mechanisms. The authors apply the proposed framework to the task of releasing counts with bounded relative error and demonstrate empirical improvements over existing methods.

Strengths

The propose framework could be useful for DP practitioners to compose traditional DP and ex-post DP mechanisms. The proposed methods provide DP guarantees subject to outcomes satisfying strict accuracy requirements.

Weaknesses

The presentation of this paper is not very clear to me. The first listed contribution seems to be the general composition theorem for ex-post differential privacy; however, it is only mentioned in the appendix. In the main text, I noticed a list of theorems, but there seems to be a lack of intuition provided, such as explanations on why these bounds are tight.

Questions

In Figure 2 left, why is the standard deviation so large? Is this expected?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

This paper is theoretical and does not have potential negative societal impact.

Reviewer xyZp6/10 · confidence 5/52023-07-27

Summary

One of the challenges any attempt to implement DP preserving techniques, is the inherent tradeoff between the level of guaranteed privacy and accuracy levels. While most of the analysis done by privacy experts focused on the perspective of maximizing accuracy subject to an upper bound on the privacy loss, many data scientists might prefer the inverse perspective, e.i., minimizing privacy loss subject to a lower bound on accuracy. This latter perspective was first formalized by [1], who referred to it as "accuracy first". They proposed a mechanism that releases a sequence of responses with increasing accuracy and decreasing privacy levels, which - somewhat surprisingly - pays only for the privacy loss of the last released response, which they analyze using "ex-post privacy". Later [2] extended these results to Gaussian noise as well using the Brownian mechanism, but since they did not rely on the classical DP notion, they lacked a composition guarantee. The authors of the current work resolved this issue, by leveraging the privacy filter toolkit, in which privacy parameters can be adaptively chosen, under the constraint that they will not exceed some pre-defined threshold. Specifically, they use a zCDP filter for the Brownian mechanism which can be used for composing several calls to the mechanism as well as other zCDP mechanisms. They use this technique to analyze a setting where the accuracy goal is defined in terms of relative error rather than additive error, and use empirical evaluation to asses the accuracy improvement achieved by applying their analysis technique. [1] K. Ligett, S. Neel, A. Roth, B. Waggoner, and S. Z. Wu. Accuracy first: Selecting a differential privacy level for accuracy constrained erm. [2] J. Whitehouse, A. Ramdas, S. Wu, and R. Rogers. Brownian noise reduction: Maximizing privacy subject to accuracy constraints.

Strengths

Providing composition guarantees for ex-post privacy is an important step towered utilizing the accuracy first perspective in the real world, which can help to further extend the usage of differential privacy techniques in the real world. The background and intuition are clearly presented, the results in the body of the paper are formally stated, and the proofs are sound.

Weaknesses

While the results presented in this paper are not just a simple application of the privacy filter technique to the ex-post privacy framework, as discussed in section 5.1, the novel results are somewhat limited in their scope, as they apply only to the zCDP privacy measure and the Brownian mechanism. The empirical evaluation is limited to a specific use case, and a single $(\epsilon, \delta)$ tuple (with $\epsilon = 10$ which is higher than what is usually considered private). In the introduction, the authors also claim to provide a basic composition theorem for arbitrary ex-post privacy mechanisms, which appears only in the supplementary material. Putting aside the fact that the result does not appear in the body of the paper, and the fact the supplementary material is not comprised of appendix but is an extended version of the paper which makes it hard to find the additions, the actual claim (Theorem 3) is somewhat hard to parse, and could probably benefit from removing the zCDP part. Small minor comment: It seems like the two references to Whitehouse et al. papers in the abstract were mixed up. The first one should've been the 22 paper and the second the 23. ======= **Edit after rebuttal discussion:** The authors have addressed my concerns.

Questions

I have no additional questions to the authors.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer dJHg2023-08-13

I stand by my overall recommendation of weak accept. The results are, I believe, novel and powerful--but for a mathematically heavy and subtle paper like this one, there are still some presentation and clerity issues that I would prefer to see resolved. I _think_ I would still appreciate some more clarity in the statement of theorem 3 (and indeed in a few other places). E.g., tracing back to understand what is the precise meaning of statistic, I wonder if I see either a small abuse of notation or a typo in definition 4.4. That is: $f$ is declared to be a function with domain $\mathcal{X}$, $T$ a function with domain $\mathcal{Y}^*$, but both are evaluated at the same variable $x$ in the definition of the BM. Definition 4.3 makes me think that there is possibly a suppressed $A$ here, but the discussion of who is a noise reduction mechanism, who is an ex-post-privacy mechanism, etc., makes me think that the BM is implicitly occurring in its own definition? It is possible that sense can be made here, but it is not obvious to me how. A super minor nit: On equation (2), in definition 4.1: I see that this is indeed not identical to epsilon-delta DP. But I think this definition remains _somewhat_ confusing--if it is novel to this paper, I would recommend reconsidering. I believe my confusion here was due to the fact that in some sense $\epsilon^*$ and $\mathcal{E}$ 'specify the same thing'--rather, that $\mathcal{E}$ has strictly _more_ information than $\epsilon^*$. IE, including $\epsilon^*$ in the definition is really just a 'highlighting' of a property of a function $\mathcal{E}$--that it is uniformly bounded above. I believe my mind went to considering equivalence with $(\epsilon, \delta)$ because I found it natural to consider $\mathcal{E}$ as a _variable_--that is, I implicitly read the definition incorrectly, assuming that $\mathcal{E}$ was not _fixed_, but rather $A$ was $(\mathcal{E}, \delta, \epsilon^*)$-ex-post-private if the claimed equation held for all $\mathcal{E}$ which are uniformly bounded above--in this case I think it would indeed be equivalent to $\epsilon, \delta$. So TLDR: I would recommend eliminating the specialization of the notion of ex-post privacy when the function $\mathcal{E}$ has a uniform upper bound, and instead prefer to track the properties of a given $\mathcal{E}$ through arguments (since as far as I can tell this tracking-through-arguments will need to happen in any case).

Authorsrebuttal2023-08-16

Thank you for the helpful comments. There is a slight abuse of notation with how we are defining the Brownian mechanism. To help clarify, we can define $BM^{(n)}(x) = (t^{(k)}, f(x) + B^{(t^{(k)})})_{k \leq n}$, and then $BM(x) = BM^{(T(x))}(x)$ where $T(x) = T(BM^{(\infty)}(x))$. Note that $T(BM^{(\infty)}(x))$ is a stopping time, so that the event $T(BM^{(\infty)}(x))= k$ only needs to know outcomes $BM^{(k-1)}(x)$. We will clarify this in the next revision. Good point about the ex-post privacy definition with three arguments. The $\epsilon^*$ is indeed a bound on the function $\mathcal{E}()$, so we can keep the original definition of ex-post privacy and set the condition on $\mathcal{E}()$ in the theorem/lemma statements.

Reviewer xyZp2023-08-14

I thank the authors for the clarifications, and have no follow-up questions.

Reviewer HnZw2023-08-15

Thank you for your response and addressing my questions. It would be great if you could add such a discussion near the theorem statements.

Reviewer njjx2023-08-18

Thank you

Thanks to the authors for answering my questions and concerns. I would like to raise my score by 1, as I think the paper is technically interesting, but the presentation might need some extra effort to polish.

Reviewer hK6N2023-08-18

Thanks for the clarifications. I've raised my score by 1; I don't want to be that one curmudgeonly reviewer holding back this paper from publication, especially as I think the results could be very useful. But like Reviewer njjx, I think that improving the presentation will be a significant under-taking.

Authorsrebuttal2023-08-19

Thank you. We have been revising our paper as we have been responding, and it is indeed already much clearer thanks to the great questions and feedback. We will continue to iterate, especially using the extra page judiciously, if the paper is accepted

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC