Private Everlasting Prediction

A private learner is trained on a sample of labeled points and generates a hypothesis that can be used for predicting the labels of newly sampled points while protecting the privacy of the training set [Kasiviswannathan et al., FOCS 2008]. Research uncovered that private learners may need to exhibit significantly higher sample complexity than non-private learners as is the case with, e.g., learning of one-dimensional threshold functions [Bun et al., FOCS 2015, Alon et al., STOC 2019]. We explore prediction as an alternative to learning. Instead of putting forward a hypothesis, a predictor answers a stream of classification queries. Earlier work has considered a private prediction model with just a single classification query [Dwork and Feldman, COLT 2018]. We observe that when answering a stream of queries, a predictor must modify the hypothesis it uses over time, and, furthermore, that it must use the queries for this modification, hence introducing potential privacy risks with respect to the queries themselves. We introduce private everlasting prediction taking into account the privacy of both the training set and the (adaptively chosen) queries made to the predictor. We then present a generic construction of private everlasting predictors in the PAC model. The sample complexity of the initial training sample in our construction is quadratic (up to polylog factors) in the VC dimension of the concept class. Our construction allows prediction for all concept classes with finite VC dimension, and in particular threshold functions with constant size initial training sample, even when considered over infinite domains, whereas it is known that the sample complexity of privately learning threshold functions must grow as a function of the domain size and hence is impossible for infinite domains.

Paper

References (29)

Scroll for more · 17 remaining

Similar papers

Peer review

Reviewer NTTD7/10 · confidence 3/52023-07-04

Summary

This work proposes the notion of private everlasting prediction. Given a training dataset, the predictor responds to a sequence of queries and privacy has to be preserved for both the training data and all queries. The authors explore the PAC learnability problem under this model and show that the sample complexity scales quadratically with the VC dimension through a generic construction from non-private learners.

Strengths

1. The authors formulate a new notion of private everlasting prediction. This is an original theoretical concept which extends the single query prediction model and has practical relevance. 2. The authors prove several interesting theoretical properties for private everlasting prediction. First it requires that the hypothesis needs to change over time. Second, the authors show that the sample complexity scales quadratically with the VC dimension of the concept class. This is a significant improvement compared to private learning which is impossible over infinite domains. 3. The writing is clear and easy to follow. Update: increased my score to seven after seeing the authors' response and other reviewer comments.

Weaknesses

1. In the algorithm GenericBBL, $\tau>1.1\times 10^10$ seems too large. Can the constant be made any smaller? 2. The algorithm is not computationally efficient. 3. In terms of writing, the authors could provide a brief overview of the proof ideas for Theorem 5.1.

Questions

In practice, some large models are pretrained on public data which can be considered non-private. Theoretically, is it relevant to consider the case when the training set $S$ is non-private and the stream of queries is private? How would the sample complexity change?

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors adequately addressed the limitations and potential negative societal impact.

Reviewer nCbQ7/10 · confidence 3/52023-07-04

Summary

The paper discusses private everlasting prediction, which extends private prediction to answer an unlimited sequence of prediction queries. The goal is to present a generic private everlasting predictor with low training sample complexity. The paper introduces definitions for everlasting prediction and everlasting differentially private prediction interfaces. It then presents a generic construction called GenericBBL for private everlasting prediction.

Strengths

1. This paper introduces a formal framework for everlasting prediction. 2. This paper provided a comprehensive privacy analysis to the proposed privacy-preserving everlasting prediction task.

Weaknesses

1. Limited applicability: The paper focuses on the theoretical aspects of private everlasting prediction and does not provide concrete practical applications or empirical evaluations. It remains to be seen how well the proposed approach translates into real-world scenarios. 2. Lack of comparative experimental analysis: The excerpt does not mention any comparison or benchmarking against existing methods or alternative approaches. Without such comparisons, it is difficult to assess the novelty or superiority of the proposed method. 3. Lack of empirical validation: The excerpt does not mention any empirical experiments, simulations, or case studies to validate the effectiveness or robustness of the proposed private everlasting predictor. It is unclear how the proposed construction performs in terms of privacy preservation, utility, and sample complexity compared to other techniques. 4. The contents shown in the supplementary materials contain a lot of redundant information compared with the paper.

Questions

What is the application scenario of the everlasting prediction?

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

It lacks experimental analysis of the proposed algorithm, which makes it hard to justify the utility of the proposed algorithm.

Reviewer xv4Q8/10 · confidence 3/52023-07-07

Summary

This paper provides an intriguing path to evading known lower bounds for differentially private PAC learning. Whereas nonprivately the sample complexity of learning is proportional to the VC dimension, the private sample complexity for (pure) DP is characterized by the representation dimension, which can be much larger. In particular, some natural classes such as threshold functions over an infinite domain (e.g. Z or R) have finite VC dimension but infinite representation dimension, so they can be learned nonprivately but not privately (even with approximate DP, by a separate result). This paper shows that this impossibility only holds for privately releasing a hypothesis and not for privately classifying samples. In particular, it shows that in the online prediction setting, for any hypothesis class there is a generic stateful algorithm with sample complexity based on the VC dimension (to be precise, quadratic in the VC dimension) that can privately answer an unbounded sequence of iid queries with the same accuracy guarantee as PAC learning, but where the state of the algorithm remains hidden and only the query labels are revealed. The algorithm is allowed to remember previous queries but must be differentially private with respect to them as well as the points of the original (labeled) training set. It also shows that statefulness is necessary to achieve this result. For general hypothesis classes the algorithm is inefficient, but it can be made efficient on important special cases, including threshold functions.

Strengths

This is a very interesting result that enhances our understanding of differentially private learning by bypassing known lower bounds. It essentially gives a separation between interactive and non-interactive private prediction, showing that interactivity reduces the private sample complexity to nearly that of nonprivate learning (as long as we only have to output predictions and not a hypothesis). The paper is quite well-written.

Weaknesses

There are two obvious limitations of the main result, that yield very interesting open problems posed in the paper: whether it's possible to come up with a similar algorithm that is computationally efficient for all hypothesis classes, and whether it's possible to improve the sample complexity from VC^2 to VC. Minor corrections: 104: points --> point 147: "upto" should be "up to"

Questions

1) Is it strictly necessary for accuracy that the query points are iid? Or could we hope for the following stronger statement: Given query points x_1, x_2, ... and a subset of indices I (not known to the algorithm) such that x_i is sampled from the distribution for i\in I but is chosen by the adversary for i\notin I, we require accuracy only for the points in I. That is, as long as the challenge queries are sampled from the distribution, we want to be robust to data poisoning using the other queries. Does this seem like it could be possible, or is it too strong? 2) Minor clarification: In the technical overview on page 3-4 it sounds like the parameters discussed are based on the standard composition bound, without using BetweenThresholds. It might be helpful to clarify whether this is the case and if we get a similar or only slightly weaker result without using BetweenThreshold, or if we truly need the tighter privacy analysis of BetweenThreshold.

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

Yes.

Reviewer 9JKU9/10 · confidence 4/52023-07-07

Summary

This work studies differentially private prediction. It has two major contributions: a) Prediction corresponds to being given an initial labeled training set, and then subsequently making predictions on other data points based on it. This paper shows that differentially private prediction can be performed on an unbounded number of queries, with strong accuracy guarantees, with the initial training set sample complexity scaling only as a polynomial in the VC dimension of the hypothesis class. Prior work either studied prediction for a small number of queries or studied the more stringent task of private PAC learning where significantly stronger lower bounds are known (PAC learning asks for the release of an entire model as opposed to predictions alone). b) Since queries correspond to data points, they are usually sensitive user information. The authors formalize a model of privacy (similar to joint differential privacy) that gives a meaningful notion of privacy for these data points, even when adversaries can choose the queries adaptively. Their algorithms satisfy this notion of privacy.

Strengths

This paper makes a valuable and somewhat surprising discovery by proving that differentially private prediction can be performed on an unbounded number of queries, with no training set sample complexity dependence on the number of queries! They leverage standard techniques such as sparse vector in clever ways to do so. They operate in rounds, and use queries themselves as data points for future rounds, calling upon techniques from the semi-supervised differentially private learning literature . Their results suggest new ways to get around lower bounds for differential privacy. Their algorithms involve reducing private prediction to non-private PAC learning, and while their techniques are not yet practical in many cases (because of time complexity), their reduction can be readily extended to any non-private learning algorithm, and hence there is lots of potential for future work to come up with ways to make their techniques more practical.

Weaknesses

One (minor) drawback is that the notion of privacy defined for the adaptively chosen queries is (necessarily) not as strong as would be nice since the label for a data point needs to depend on the data point to achieve any reasonable notion of accuracy (since otherwise you can only do something like randomized response).

Questions

1) The reasoning for the use of LabelBoost didnt entirely make sense to me- the stated reason was that the predictions provided during a round are not consistent with any concept in the hypothesis class. However, it’s not immediately clear to me that this is a problem- for example, using agnostic (non-private) learning algorithms may get around this issue (at the expense of worse sample complexity dependence on the accuracy parameter). More explanation of this would be useful. 2) The comment about their model of prediction, in principle, allowing memorization was interesting. It’s not clear that this is true- it seems like memorizing something in the training set would affect future predictions which would affect the adversary’s views. Would love more clarification on this. 3) Is there a simple (eps, 0)- version of this algorithm? It’s not clear to me that approx DP is necessary (sparse vector by itself is pure DP, though a variant is used in this paper). 4) The diagrams in the privacy proofs are a little too convoluted to be useful to the reader. I don’t have great suggestions on how to simplify them, but the privacy proofs by themselves are relatively simple, and I’m not sure the diagrams make them simpler to understand. 5) Some minor comments: a) In Algorithm LabelBoost step 4, ‘choose’ based on exponential mechanism is not well defined since the score function is not specified (it can be inferred from context that it’s the negative of empirical risk) but specifying would be good. b) In Claim E.4 don’t you need to account for the fact that the function labeling the sample is not the real function, but rather the one chosen by LabelBoost? Specifically, Claim E.3 is about a function with 0 sample error, whereas LabelBoost will incur some error, and hence I think you want a version of claim E.3 where sample error is bounded by alpha. c) On page 8, where it says ‘using standard privacy amplification arguments, BetweenThresholds can be modified to allow for c times of outputting bot before halting)’, I think you mean privacy composition.

Rating

9: Very Strong Accept: Technically flawless paper with groundbreaking impact on at least one area of AI/ML and excellent impact on multiple areas of AI/ML, with flawless 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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Limitations adequately addressed.

Reviewer NTTD2023-08-14

Thank you for your resonse

Thank you for your response. I believe the new concept is very meaningful and could inspire a new line of research. I will increase my score to 7.

Reviewer nCbQ2023-08-15

Thank you for your response

Thank you for the explanation! It addresses my concerns and I agree that the new concept is very meaningful. I would like to increase my score to a 7.

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC