Private estimation algorithms for stochastic block models and mixture models

We introduce general tools for designing efficient private estimation algorithms, in the high-dimensional settings, whose statistical guarantees almost match those of the best known non-private algorithms. To illustrate our techniques, we consider two problems: recovery of stochastic block models and learning mixtures of spherical Gaussians. For the former, we present the first efficient $(\epsilon, \delta)$-differentially private algorithm for both weak recovery and exact recovery. Previously known algorithms achieving comparable guarantees required quasi-polynomial time. For the latter, we design an $(\epsilon, \delta)$-differentially private algorithm that recovers the centers of the $k$-mixture when the minimum separation is at least $ O(k^{1/t}\sqrt{t})$. For all choices of $t$, this algorithm requires sample complexity $n\geq k^{O(1)}d^{O(t)}$ and time complexity $(nd)^{O(t)}$. Prior work required minimum separation at least $O(\sqrt{k})$ as well as an explicit upper bound on the Euclidean norm of the centers.

Paper

References (66)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer cXQY8/10 · confidence 4/52023-07-02

Summary

In this paper, the authors investigate graph clustering with the Stochastic Block Model (SBM) and metric clustering with the Gaussian Mixture Model (GMM) within a novel privacy-preserving framework. They propose differentially private (DP) algorithms for both problems, which offer significantly improved performance guarantees compared to state-of-the-art DP algorithms. Specifically, the authors present efficient $(\epsilon,\delta)$-differentially private SBM algorithms for both weak recovery and exact recovery. These algorithms operate in polynomial time, outperforming the previous quasi-polynomial time DP algorithms. Additionally, they provide a lower bound that demonstrates the optimality of the privacy guarantees from an information-theoretic perspective. Furthermore, the authors develop private algorithms for learning the mixture of $k$ spherical Gaussians under the condition that the minimum separation is at least $O(k^{1/t}\sqrt{t})$. Notably, these algorithms do not require an additional additive term of $\Omega(\sqrt{\log n})$ in the minimum separation or an explicit upper bound on the Euclidean norm of the centers, which were assumed in prior works. The proposed algorithms are built upon a simple yet powerful connection between robustness and privacy. The authors formalize this connection as a framework for transforming robust estimation algorithms into private algorithms. They show that if two strongly convex functions over constrained sets, where both the function and the set may depend on the input, are point-wise close, their minimizers are also close. Conversely, they establish that projections of points that are close to each other onto convex sets that are point-wise close must also be close.

Strengths

The paper is well-written and addresses fundamental clustering problems, including SBM and GMM, providing novel differentially private algorithms that make remarkable advancements in the field. The lower bound also demonstrates the optimality of the privacy guarantees of the SBM algorithm. Moreover, their techniques of translating robust estimation algorithms to DP algorithms seem readily extended to some other estimation algorithms. Though the ideas are intuitively simple, the formal connections and applications to clustering algorithms are highly nontrivial.

Weaknesses

There is no experiment in the paper (but it is a strong theory paper).

Questions

No.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

It might be useful to briefly discuss why the current framework does not achieve $(\log k)^{c+1/2}$ separation in polynomial time.

Reviewer SDAx8/10 · confidence 3/52023-07-06

Summary

This paper introduces a general framework for designing efficient private estimation algorithms from robust non-private algorithms. Their approach, based on a certain stability analysis for convex optimization, is applied to weak and exact recovery in the stochastic block model (SBM) and for learning mixtures of spherical Gaussians. They obtain the first efficient $(\varepsilon,\delta)$-private algorithms for estimation in the first setting (with a new lower bound suggesting that there is little room for improvement), as well as one for the second setting which allows for far lower separation between centers than before.

Strengths

- The approach introduced in this paper appears very useful and flexible, and I anticipate future work building off this approach to address a variety of high-dimensional private estimation tasks - They make significant progress on two important problems in private estimation - Presentation is strong, prior work and guarantees are clearly delineated, their results are clearly stated, and, while there is a bit of technicality in the second application, the authors build up to it nicely with the first SBM results - The Gaussian mixture application requires some clever sum-of-squares analysis. More broadly, there is certainly technical depth to this work beyond the general framework

Weaknesses

- I sometimes lost track of the role of different parameters. I think some may be obvious to those more familiar with this subfield of privacy, and the many special cases mentioned in the discussion were helpful. In any extended version of this work, I would suggest adding more discussion of terms appearing in bounds from the classic (non-private) settings. - The choice of objective function in the Gaussian mixture application wasn't clear to me - where does the all ones matrix come from? - A fair bit of technical detail was pushed to the supplement, which was a bit tricky to parse. In particular, while the high level approach was clear to me, I was unable to verify correctness for the Gaussian mixture proofs (though I mostly blame this on my unfamiliarity with the SoS program they consider) Minor nits: - On lines 207-208, it was unclear to me why $k \geq \sqrt{\log d}$ should be considered a high-dimensional regime - On line 272, the objective function $f$ is not defined (though clear enough from context) - "the basic" repeated on line 360

Questions

In addition to those above: - Are the authors aware of their sensitivity analysis appearing in any previous convex programming literature? It is simple enough that I am a bit surprised it was not previously known. For instance, if their Condition (ii) was replaced with a Hausdorff distance bound between K(Y) and K(Y'), I think this is essentially folklore (though of course Condition (ii) is more sophisticated than this).

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed 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

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors were clear about all assumptions and I don't anticipate any negative impacts.

Reviewer NQm57/10 · confidence 3/52023-07-10

Summary

The paper studies private parameter estimation in certain high-dimensional statistical inference tasks. In particular, community detection in stochastic block models (where the privacy is with respect to edge changes) and estimation of unknown means in isotropic Gaussian mixture models. Although these problems have been studied in the literature, the paper provides new algorithms with improved sample and computational complexities, which closely match that of non-private estimation. Although seemingly disjoint, the paper's results follow a common methodology of privatizing the existing non-private algorithms that rely on (strongly) convex optimization.

Strengths

The paper proposes a generic privatization procedure relying on strongly convex optimization, and then the paper applies it to two problems of interest, leading to improved sample and computational complexities. The paper's results are novel and interesting.

Weaknesses

I have the following comments: 1. The paper's writing needs to be improved and it seems that the submitted version is missing many details. For example, the notation for the indicator function is not displayed due to LaTeX error and many algorithmic environments are referenced as theorems by cleveref. 2. The comparison with work of Kothari-Manurangsi-Velingker is still not clear. Perhaps the paper should begin by explaining the similarities and then the differences. Please also see the comments below.

Questions

# Major Comments 1. I suggest adding the notion of privacy (node/sample differential privacy) in the theorem statements. 2. Page, Line 45: "However, for the more challenging task of graph clustering, node-DP is sometimes impossible to achieve." I do not follow. For the SBM, I believe that node-DP is possible to achieve since the information is not concentrated on a single node. 3. Footnote 11: I suggest explicitly adding the notion of privacy in the theorem statement (see the comment above as well). 4. Line 259: Where is the first step (i) from Line 253 used in the result here? Does it not follow from (ii) alone? Moreover, what if $K(Y)$ is empty? 5. Line 270, The convex program for SBM. The program from the citation [24] is different from the one displayed here. Especially, it is linear in X and not quadratic. Please clarify. # Other Comments 1. Page 2, Line 49: What is meant by "non-robust setting" 3. Page 2, Line 64: I believe there are more representative papers for outlier robust estimation. 4. Page 3, Model 1.1: Please give intuition about the parameters in the model, for example, d is the average degree. 5. Theorem 1.3: Please add the weak recovery guarantee in the theorem statement. 6. Line 266: Describe explicitly the relation between the input and Y 3. Lines 182-188: I found the whole paragraph confusing. I suggest rephrasing. 4. Theorem 1.6: Why does $k$ need to be large? 6. Line 210: Put $t^*$ in the runtime or say that it is constant. 7. Line 250-251 Please add citations here "Our sensitivity bound is simple, yet it generalizes previously known bounds for strongly convex optimization problems" 8. Line 277: This was confusing and what is the norm $\|\cdot\|_1?$. Please rephrase and explain that the function $f$ does not contain the quadratic in $Y$. Although it does not affect the optimization problem, but it is important for the sensitivity analysis. 9. Line 312: "While this program suggests a path to recover the first property,..." Is it even feasible? 10. Line 321: Briefly explain that it is so by setting $z_i$ to zero. 11. Line 341-347: This was confusing for me. What histograms are being referred to here? I understand the space might have been an issue originally but the accepted version can be longer.

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

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

4 excellent

Presentation

2 fair

Contribution

4 excellent

Limitations

N/A

Reviewer GVoS7/10 · confidence 4/52023-08-01

Summary

The paper considers the problems of weak and strong recovery for stochastic block models, as well as the problem of estimating the means of Gaussian Mixture Models with sphrerical Gaussians, under the constraint of differential privacy. Their contributions include: * A computationally efficient algorithm that performs strong recovery for the stochastic block model under $(\varepsilon, \delta)$-edge privacy. * A computationally inefficient algorithm that performs both weak and strong recovery for the stochastic block model under $\varepsilon$-edge privacy, as well as a matching information-theoretic lower bound. * A computationally efficient algorithm for estimating the means of mixtures of spherical Gaussians under $(\varepsilon, \delta)$-DP. The algorithms are all derived by applying a general framework that is developed as part of the work which involves privatizing the solutions of strongly convex optimization programs. Additionally, the results given in the paper improve upon the prior work of Seif et. al. and Kamath et. al. on SBMs and GMMs, respectively.

Strengths

As mentioned in my summary, the paper provides a general framework for privatizing the solutions of convex programs, which they apply in a multitude of settings. I consider this to be a significant contribution in the private statistics literature. In addition to that, the paper is very clearly written and the description of the techniques if quite accessible.

Weaknesses

No concrete weaknesses.

Questions

There has been recent work by Arbas et. al. on private estimation of GMMs (see https://arxiv.org/abs/2303.04288). I read both works and understand that they go in different directions (meaning that the work of Arbas et. al. doesn't supersede the present paper). However, unless I missed something the paper is not mentioned at all in the related work section (or anywhere else in the submitted pdfs), which feels like an important ommission. Thus, it'd be great if the authors could provide a comparison of their work with that paper (and also include it in their discussion of related work).

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

This is a purely theoretical work with no obvious societal impact.

Reviewer SDAx2023-08-10

Thanks for providing these details. These sufficiently address the points I raised, and I intend to keep my score.

Reviewer NQm52023-08-10

I thank the authors for their detailed response. Could you please also explain the similarities and dissimilarities with the Kothari-Manurangsi-Velingker paper in detail?

Authorsrebuttal2023-08-14

Comparison with Kothari-Manurangsi-Velingker

Both Kothari-Manurangsi-Velingker (KMV) and our work obtained private algorithms for high-dimensional statistical estimation problems by privatizing strongly convex programs, more specifically, sum-of-squares (SoS) programs. On a high level, we both first solve an SoS program, and then add Gaussian noise to the solution, and return the result. The main difference between KMV and our work lies in how we choose the SoS program. Such a difference saves us a computationally-expensive but necessary step for KMV. For the problem of robust moment estimation, KMV considered the canonical SoS program [1] which contains a minimum cardinality constraint. In robust statistics literature, such minimum cardinality constraints are added to ensure good utility, without privacy concerns [1]. However, from the perspective of privacy, such a constraint is problematic, since there will always exist two adjacent input datasets such that the constraints are satisfiable for one but not for the other. KMV and us resolve this privacy issue in different ways as follows: - KMV uses an exponential mechanism to pick the lower bound of the minimum cardinality constraint. This step also ensures that solutions to the resulting SoS program will have low sensitivity. - In contrast, we simply drop the minimum cardinality constraint. Then the resulting SoS program is always feasible for any input dataset! However, the resulting SoS program will not have good utility if we do nothing else. To have good utility, we need to pick an appropriate objective function. For example, in Gaussian mixture models, we dropped the constraint that $\sum_{j}W_{ij}$ should be lower bounded and turned to minimize $||W-J||_F^2$ (where $J$ is the all-one matrix). Our approach provides some advantages over KMV. From a purely algorithmic point of view, the exponential mechanism in KMV requires computing $O(n)$ scores. Computing each of these scores requires solving a large semidefinite program. In contrast, our algorithm completely avoids this step. This difference can provide significant improvement in the running time. From the point of view of algorithm analysis, proving that the exponential mechanism in KMV works requires several steps: 1) defining a (clever) score function, 2) bounding the sensitivity of this score function and, 3) showing existence of a large range of parameters with high score. Our approach completely bypasses these steps by directly bounding the sensitivity of the program via the key lemma, hence significantly simplifying the proofs. [1] Kothari, Pravesh K., and David Steurer. "Outlier-robust moment-estimation via sum-of-squares." _arXiv preprint arXiv:1711.11581_ (2017).

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC