Optimal Unbiased Randomizers for Regression with Label Differential Privacy

We propose a new family of label randomizers for training regression models under the constraint of label differential privacy (DP). In particular, we leverage the trade-offs between bias and variance to construct better label randomizers depending on a privately estimated prior distribution over the labels. We demonstrate that these randomizers achieve state-of-the-art privacy-utility trade-offs on several datasets, highlighting the importance of reducing bias when training neural networks with label DP. We also provide theoretical results shedding light on the structural properties of the optimal unbiased randomizers.

Paper

Similar papers

Peer review

Reviewer oAvy7/10 · confidence 4/52023-06-30

Summary

This paper investigates the bias of the state-of-the-art label-differential privacy (label-DP) mechanism proposed by Ghazi et al. (2023) and proposes bias-corrected randomizers achieved through minimizing of a constrained linear programming approach. The proposed label-DP mechanisms demonstrate lower mean squared errors empirically by carefully tuning the hyperparameters in deep learning.

Strengths

1. The main result of this paper is intriguing and original. The paper provides a compelling example illustrating the bias present in the state-of-the-art label-DP mechanism and proposes a new mechanism to address this bias. Additionally, the authors thoroughly investigate the empirical performance of the proposed algorithm and show that it outperforms the state-of-the-art. 2. This paper is well-written. The authors effectively communicate their contributions and main findings in a clear and concise manner. Furthermore, the organization of the paper is well-structured.

Weaknesses

1. The proposed algorithm introduces extra computational cost. In order to obtain the unbiased randomizer, a constrained linear programming (LP) problem needs to be solved, which introduces extra computational cost. Although the authors demonstrate the feasibility of the LP, the term "feasible" left me uncertain, and it would be more informative to explicitly state the computational complexity involved. 2. The authors should consider the potential privacy implications of hyperparameter tuning. In the experimental section, they mention the careful tuning of hyperparameters. However, if the validation set used for this tuning procedure is not label DP, there is a risk of privacy leakage.

Questions

See the weakness part.

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

Yes

Reviewer YHYY7/10 · confidence 4/52023-07-06

Summary

This paper proposes a differentially private algorithm for regression problems. The algorithm protects the privacy of the labels ("label DP"), in contrast to the entire example. The canonical application of this is digital advertising, where the label might be transaction data from a separate website. Furthermore, the algorithm operates in what [GKK+23] term the "feature-oblivious" model, where a private algorithm, operating solely on the labels, sends a message to the "features party," who then uses these together to learn a model. [GKK+23] studied a label randomizer that aims to minimize the difference between the true and noisy labels. This work studies a unbiased randomizers. In addition to experiments, the paper provides some theoretical evidence that unbiased estimators, even if they have high variance, may be superior. The private algorithm (which, again, only has access to the labels) proceeds in two steps. First, it uses a private histogram to estimate a prior over the labels. It then solves a linear program, the output of which is our label randomizer. The LP minimizes the expected difference between the true and noisy labels, subject to privacy, unbiasedness, and normalization constraints. An important technical question is the randomizer's output space. Let $\mathcal{Y}\subseteq \mathbb{R}$ be the label space and $\hat{\mathcal{Y}}\subseteq \mathbb{R}$ the set of possible outputs of the randomizer. The solution to the LP has $|\mathcal{Y}|\times|\hat{\mathcal{Y}}|$ entries, representing a probability distribution over $\hat{\mathcal{Y}}$ for each entry in $\mathcal{Y}$. Thus it is clear that this approach demands both sets be finite (and of modest size, to solve the LP). What is less clear, due to the unbiasedness constraint, is what $\hat{\mathcal{Y}}$ should be. Note that if $\min \mathcal{Y} =0 $ and $\hat{\mathcal{Y}}=\mathcal{Y}$, then unbiasedness demands we map $0\to 0$ with probability 1, violating privacy. Thus the endpoints of $\hat{\mathcal{Y}}$ must exceed those of $\mathcal{Y}$. The authors set the endpoints of $\hat{\mathcal{Y}}$ so that (provably) the LP is feasible. They "fill in" the rest of $\hat{\mathcal{Y}}$ with a grid, as finely spaced as their computational constraints allow. For a fixed $\hat{\mathcal{Y}}$ and prior on $\mathcal{Y}$, the LP finds the unbiased private randomizer with the lowest expected difference (or loss) in labels. Experiments demonstrate improvements over prior work on three data sets. I noticed that parts of the submission (introducing the label DP recipe, describing data sets, and reviewing related work) had substantial overlap with text from GKK+23. After discussion with my chair, I have not used this information in my evaluation of the paper.

Strengths

Without unbiasedness constraints, GKK+23 find the best randomizer is of a form they call "RR-on-Bins." The authors nicely sum up their core innovation: "the addition of an unbiasedness constraint to the linear program leads to solutions to the LP that (i) are not RR-on-Bins solutions, (ii) can have substantially higher variance than RR-on-Bins, and (iii) nevertheless have a much lower train and test error due to the reduction in bias." This is a clear observation that allows them to move past prior work. I buy that this is a practical problem, that the feature-oblivious model is useful. Their approach does seem to significantly improve on that of GKK+23 on the data sets they tested. Their experimental results are complemented by their formal claims (Theorem 6 and Proposition 7) about the output label space of randomizers. With minor exceptions, I found the presentation very clear throughout.

Weaknesses

In light of the work of GKK+23, this paper has limited novelty. Originality is also a serious weakness: GKK+23 point out that "our noising mechanism typically introduces a bias; mitigating it, say by adding unbiasedness constraints into the LP, is an interesting direction." The title says the work finds "optimal" unbiased randomizers, but that optimality only holds once the choice of output space is selected. Theorem 6 (there is an optimal unbiased randomizer with at most $2|\mathcal{Y}|$ labels) and Proposition 7 (selecting endpoints to ensure feasibility) are good steps, but do not get us all the way there. As far as I can tell, it remains possible that some other scheme generates better unbiased randomizers, at least for some priors. The paper's focus is on feature-oblivious algorithms, but I would have liked to see a comparison with other label-DP algorithms that solve this problem. How much are we giving up by moving to feature-oblivious?

Questions

Is there evidence (even informal) that this discretization approach is close to optimal? For instance, can we rule out that the LP's objective is highly sensitive to the choice of endpoints (in the regime considered)? Can you briefly sketch out what you feel are the key innovations beyond GKK+23? Estimating the prior requires some privacy budget for each bin. Do you expect that, as $|\mathcal{Y}|$ grows, another approach will perform better than yours?

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

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

The algorithm operates in the feature-oblivious model; other label-DP algorithms might perform better in different settings. The algorithm assumes prior knowledge of $\mathcal{Y}$; it is not always clear where this information comes from.

Reviewer DWLv6/10 · confidence 3/52023-07-07

Summary

The paper studies the regression problems under label differential privacy (DP). It proposes a novel randomizer that generates high-quality DP labels which can be used to train a regressor. The proposed randomized mechanism is sound and the experiments on various benchmark datasets and different privacy budgets demonstrated the improvements over the baselines. Overall I think the contribution of the paper is significant, but the proposed mechanism seems to be limited to the case where the labels are discrete (not continuous). Based on that I recommend a weak accept for this paper.

Strengths

1. The contribution is significant and the idea is novel. The paper proposes a novel randomized mechanism that can generate unbiased DP labels which helps to train an unbiased regressor. 2. The paper conducts extensive experiments on different benchmark datasets and different epsilon and shows that a noticeable reduction of mean squared error when using the proposed mechanism over the previous works.

Weaknesses

1. The paper focuses on optimal mechanism for epsilon-DP and this topic is pretty studied in some previous work such as the stair-case mechanism. It would be great if there is any discussion about extension to (epsilon, delta)-DP. 2. The settings are limited to the discrete (but not continuous) choice of labels which only applies for ordinal regression. 3. The authors did not attach the codes so it might be hard to reproduce the experiments.

Questions

Can the authors explain more why the proposed mechanism is better than the optimal stair-case mechanism for example if we look at Figure 4.a , especially even when epsilon is large? I used the staircase mechanism before and this turned out to be better than many baseline mechanisms when epsilon is large enough.

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

3 good

Contribution

3 good

Limitations

The paper discusses clearly the limitations of their proposed mechanism compared to the baselines in the Conclusion section.

Reviewer Vhiv7/10 · confidence 3/52023-07-27

Summary

This paper proposes a new family of DP label randomizers for regression models. They show that these randomizers improve the MSE of the training set at the expense of the noisy label loss, indicating an alternate bias-variance tradeoff to other similar works.

Strengths

The strength of this paper is the theoretical and empirical evidence of the randomizer. It is evident that it outperforms the unbiased randomizers by significant margins in the test error, but this comes at the cost of the noisy label loss.

Weaknesses

However, it would have been helpful to include what this tradeoff means practically and its implications.

Questions

Overall, I think the paper is well-written and has enough experiments to validate its conjectures. The novelty of the paper is in its theoretical results and improved guarantees. Some points/comments that could be improved are: $\hat{y}$, the noisy label, is never defined in Section 3? The end of Section 3, particularly the observations for $\epsilon$ are a bit unwieldy to read -- this would probably be better presented with a table or figure Why is the computational complexity not compared between the unbiased and biased randomizers? This would help understand what tradeoffs of this method are It would be nice to see an analysis of why the method works better on one dataset versus the other In Section 4.3, it is mentioned that for smaller $\epsilon$’s one can see the standard error increase -- is this only true for these types of algorithms or simpler models as well?

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

Yes

Reviewer YHYY2023-08-12

Thanks for the detailed response. Your elaboration on the discretization makes me feel much better about the approach. I will increase my score to a 7. On estimating the prior: of course you are correct, apologies for the mistake. If $|\mathcal{Y}|$ is large relative to the number of samples, our estimate of the prior might be poor. I doubt this is of much concern; as you mention elsewhere, a different discretization or grouping of labels would likely work well. (No need to reply to this comment.)

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC