Attack-Aware Noise Calibration for Differential Privacy

Differential privacy (DP) is a widely used approach for mitigating privacy risks when training machine learning models on sensitive data. DP mechanisms add noise during training to limit the risk of information leakage. The scale of the added noise is critical, as it determines the trade-off between privacy and utility. The standard practice is to select the noise scale to satisfy a given privacy budget $\varepsilon$. This privacy budget is in turn interpreted in terms of operational attack risks, such as accuracy, sensitivity, and specificity of inference attacks aimed to recover information about the training data records. We show that first calibrating the noise scale to a privacy budget $\varepsilon$, and then translating {\epsilon} to attack risk leads to overly conservative risk assessments and unnecessarily low utility. Instead, we propose methods to directly calibrate the noise scale to a desired attack risk level, bypassing the step of choosing $\varepsilon$. For a given notion of attack risk, our approach significantly decreases noise scale, leading to increased utility at the same level of privacy. We empirically demonstrate that calibrating noise to attack sensitivity/specificity, rather than $\varepsilon$, when training privacy-preserving ML models substantially improves model accuracy for the same risk level. Our work provides a principled and practical way to improve the utility of privacy-preserving ML without compromising on privacy. The code is available at https://github.com/Felipe-Gomez/riskcal

Paper

References (66)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer tsqL5/10 · confidence 4/52024-06-17

Summary

The paper proposes calibrating the noise in privacy mechanisms directly to MIA success metrics like advantage or TPR at low FPR instead of calibrating to a given $(\epsilon, \delta)$-bound and converting that to bounds on MIA success. The paper develops an algorithms for direct calibration for various MIA success metrics, most importantly TPR at low FPR. The algorithms use a discrete privacy loss random variable that is output by an existing privacy accountant. The paper compares the calibration methods by comparing the noise variances, and finds that calibrating directly significantly reduces the amount of noise required. The paper also does a small experiment showing that the reduction in noise translates to improved accuracy for machine learning.

Strengths

The paper is mostly easy to read and understand, given the amount of theory. The case for calibrating directly to membership inference success metrics is made very well, and the algorithm for the calibration seems very practical.

Weaknesses

Some parts of the proof of Theorem 3.4, the main theorem of the work, are not fully clear, and could be missing details needed to make them fully correct. I've collected these points in the Questions section. I think the issues can be fixed, and the theorem is likely true, but these should be addressed before the paper is accepted. Figure 5 shows that the $\delta'$ corresponding to the attack TPR/FPR calibration is very large compared to the standard $\delta$. The standard $\delta$ is chosen to be small, since the mechanism that randomly outputs one individual's data is $(0, 1/n)$-DP with $n$ datapoints. I think the paper should discuss how this mechanism behaves with the TPR/FPR calibration. My quick calculation suggests that with the optimal attack for this mechanism, $FNR = (1 - 1/n) \cdot (1 - FPR)$, which would give a large FNR at small FPR, suggesting that the mechanism is very private, even though it obviously isn't. Requiring the FNR bound to be large to fix this could reduce the apparent advantage that the paper's results suggest, since the paper always considers smaller FNR bounds. Minor points: - Line 97: the domain and range of $\epsilon_\omega$ are the wrong way around. - The neighbourhood relation is inconsistent: Section 2.1 describes substitute neighbourhood, while Section 2.2 describes add/remove neighbourhood. - TPR does not necessarily need to be high for a MIA to be a relevant threat. For example, if TPR = 1% at FPR = 0.01%, the attack can very accurately identify the presence of a small number of individuals, which violates the privacy of those individuals. - The results in Figure 1 could have uncertainty estimates. - Restating the theorems before their proofs and having the proofs in the same order as the main text would make the proofs much easier to read. - Line 649: $\phi$ outputs the probability that the sample came from $Q$.

Questions

Issues with Theorem 3.4: - In the reasoning around lines 673-675, why is it not possible that $\alpha$ just happens to be one of the $k+1$ possible values of Eq. (42)? On line 659, it is said that this reasoning should work for all $\alpha$. This case is considered explicitly later, but only after this line of reasoning is concluded. - Doesn't the conclusion that there can be multiple ways values of $\gamma$ and $\tau$ that satisfy the constraint on line 695 imply that the choice of the optimal test $\phi_{\tau,\gamma}$ is not unique, though the FPR and FNR are unique? - Why is the test that is found on lines 690-698 optimal? - It is not clear whether Theorem 3.4 claims "for all $\tau$ and $\gamma$" or "for some $\tau$ and $\gamma$". Minor points: - I don't think Algorithm 1, line 2 works with $\alpha = 1$, the strict inequality is never satisfied. Though this should be easy to fix. - Line 164: should this have $\alpha^\star$ instead of $\alpha$? - Are $P$ and $Q$ the correct way around in Eq. (1)? They seem to be the other way around in Gopi et al. [2021]. - Should the support of $Y$ in line 666 be $y_1, \dotsc, y_{l+1}$ instead of $x_1, \dotsc, x_{l+1}$? - Are the image classification results in Figure 1 comparable between standard and attack risk calibrations, since you had to use RDP due to data augmentations with the standard accountant? - What type of subsampling (Poisson, sampling with/without replacement) and which neighbourhood relation did you use in the DP-SGD experiments? The fixed batch sizes suggest sampling without replacement, in which case the issues that Lebeda et al. (2024) have recently raised with privacy accounting for sampling without replacement might affect your results. Reference: - Lebeda et al. "Avoiding Pitfalls for Privacy Accounting of Subsampled Mechanisms under Composition" (2024) arXiv:2405.20769

Rating

5

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

The paper discusses limitations adequately.

Authorsrebuttal2024-08-07

Additional response regarding the minor points

> The neighbourhood relation is inconsistent: Section 2.1 describes substitute neighbourhood, while Section 2.2 describes add/remove neighbourhood. Thank you for pointing this out! Our analyses are independent of the choice of the neighborhood relation, and our experiments are done with the add-remove relation. We will remove the substitution relation in Section 2.1 paragraph “Setup and notation”, mention that we use the add-remove relation in our experiments in Section 2.1, paragraph “DP-SGD”. Moreover, in Section 2.2 we will emphasize that our results are not tied to the add/remove relation. > TPR does not necessarily need to be high for a MIA to be a relevant threat. For example, if TPR = 1% at FPR = 0.01%, the attack can very accurately identify the presence of a small number of individuals, which violates the privacy of those individuals. This is a great point! Ultimately, defining acceptable thresholds on TPR/FPR, as we mention in the Concluding Remarks, is an open problem. We note that acceptable thresholds can be made in a context and application specific manner. For example, from the computer security literature (see, e.g., [Wang, 2018](https://arxiv.org/abs/1802.05409)), we know that if the prior probability of membership is lower than the attack’s FPR, then the majority of the attack's positive predictions will be incorrect, even if TPR is 100% (Wang 2018, p.2; formally, prior-aware positive predictive value, ppv = P(member | attack predicts 'member') can be low even if TPR = P(attack predicts 'member' | member) is high, as it depends on the base rate P(member) and is more significantly influenced by FPR than TPR). Our framework fits nicely with this literature. If in a particular context we have a guess for the adversary’s prior probability, this can be used to inform an acceptable TPR/FPR thresholds. We will expand the short discussion in the Concluding Remarks on this limitation as follows: “We leave open the question on how to choose the target FPR $\alpha^\star$, e.g., whether standard significance levels in sciences such as $\alpha^\star = 0.05$ are compatible with data protection regulation, **as well as what are the acceptable attack TPR levels for a given FPR**. Future work is needed to develop concrete guidance on the choice of target FPR **and TPR** informed by legal and practical constraints.” > The results in Figure 1 could have uncertainty estimates. Thank you for the suggestion. In the new experimental demonstration on private histogram release, described in the [general response](https://openreview.net/forum?id=hOcsUrOY0D&noteId=4H6Ec2IhLz), we added uncertainty regions for utility measurements done over 100 mechanism runs with different random seeds. > Restating the theorems before their proofs and having the proofs in the same order as the main text would make the proofs much easier to read. We will implement this in the final version! --- > Are $P$ and $Q$ the correct way around in Eq. (1)? They seem to be the other way around in Gopi et al. [2021]. $P$ and $Q$ are the correct way in Gopi et al. We will fix this in the final version. > Are the image classification results in Figure 1 comparable between standard and attack risk calibrations, since you had to use RDP due to data augmentations with the standard accountant? Thank you for raising this point. We indeed agree that this comparison is somewhat unfair. We now reanalyzed the training pipeline from Tramer & Boneh, 2021 using Doroshenko et al. accounting so that the methods are exactly comparable. We provide the updated figure in the PDF attached in the [general response](https://openreview.net/forum?id=hOcsUrOY0D&noteId=4H6Ec2IhLz). Indeed, the TPR values obtained with standard calibration and tight accounting are somewhat lower than with RDP accounting, especially in the small $\alpha$ regime, but the general trend remains: attack-aware calibration significantly increases task accuracy at the same risk level. > What type of subsampling and which neighbourhood relation did you use in the DP-SGD experiments? We use the add-remove relationship in our experiments, as it is standard with modern accountants. We use Poisson sampling. The batch size in the experimental details is supposed to mean the _expected batch size._ To avoid confusion, we will clearly specify the neighborhood relation and write in the format “subsampling rate $p = 0.003801$ (corresponding to the expected batch size of 256)”. > Line 97: the domain and range of epsilon_omega are the wrong way around. Line 164: should this have $\alpha^\star$ instead of $\alpha$? Should the support of $Y$ in line 666 be $\\{y_1, …, y_{l+1}\\}$ instead of $\\{x_1, …, x_{l+1}\\}$? Line 649: $\phi$ outputs the probability that the sample came from $Q$. I don't think Algorithm 1, line 2 works with $\alpha = 1$, the strict inequality is never satisfied. Though this should be easy to fix. That’s right, thank you so much for spotting these! We will fix them.

Reviewer tsqL2024-08-08

Thank you for the comprehensive response. You addressed most of my concerns, most importantly those regarding the proof of Theorem 3.4. Regarding my point about catastrophic failures, it seems that I didn't make my concern quite clear in the initial review. You are correct in saying that many mechanism do not actually allow the catastrophic privacy failure, so it may not be necessary to set a conservative TPR@FPR bound to account for them. However, the same reasoning can be used to argue that we do not need $\delta \ll \frac{1}{n}$. As a result, the utility increase of TPR@FPR calibration over $(\epsilon, \delta)$-calibration seems to come from using this argument for one definition, but not the other. Reading the paper again, I noticed that you have alluded to this in the caption of Figure 5, but this point is important enough to discuss in the main text. Currently, the Abstract and Introduction give the impression that the utility benefit is due to some intrinsic difference between $(\epsilon, \delta$) and TPR@FPR bounds, and not simply a consequence of what values are typically considered acceptable for the bounds. On the other hand, I recognise that calibrating to attack risk is useful, and your method for this calibration is much easier to use than first finding the optimal $(\epsilon, \delta)$-bound corresponding to the attack risk, and then calibrating the mechanism. As a result, the paper makes a valuable contribution, and I'm increasing my score accordingly.

Authorsrebuttal2024-08-08

> You are correct in saying that many mechanism do not actually allow the catastrophic privacy failure, so it may not be necessary to set a conservative TPR@FPR bound to account for them … However, the same reasoning can be used to argue that we do not need $\delta \ll \frac{1}{n}$ … Currently, the Abstract and Introduction give the impression that the utility benefit is due to some intrinsic difference between $(\varepsilon, \delta)$ and TPR@FPR bounds, and not simply a consequence of what values are typically considered acceptable for the bounds. Thank you for raising this point. Just to clarify, as long as the TPR@FPR bounds are computed using _the privacy curve $\varepsilon(\delta)$ (using the approaches in Appendix A) / PLRVs $(X, Y)$ (using Algorithm 1)_ and not based on a single $(\varepsilon, \delta)$-point (using Eq. 5), then indeed there is no intrinsic difference. There **is**, however, an intrinsic difference in calibrating to a target TPR@FPR using Algorithm 1 vs. calibrating to a target TPR@FPR with _a fixed delta_ (standard calibration), i.e., the difference between the two curves in Figure 4. We appreciate your time and willingness to revise your review, and we are happy to answer any other questions you may have. Your input has been very valuable!

Reviewer pcXQ6/10 · confidence 3/52024-07-10

Summary

This paper proposes new methods for calibrating noise in differentially private learning to achieve a given level of operational privacy risk, specifically focusing on the advantage and FNR/FPR of membership inference attacks. The methods reduce the noise scale compared to the standard two-step procedure (first converting to a privacy budget, then converting to a privacy assessment).

Strengths

1. The paper addresses an important practical problem in the privacy regime with significant theoretical contributions. 2. It is well-organized and easy to follow.

Weaknesses

1. I suggest the authors include more downstream tasks and utility metrics to further demonstrate the effectiveness of the theoretical results. 2. High-level intuitions for different variables in Theorem 3.4, such as $\alpha(\tau, \gamma)$ and $\beta(\tau, \gamma)$, would be beneficial.

Questions

1. Will choosing a discrete-valued dominating pair of the mechanism be a sub-optimal choice for advantage calibration? 2. How is the discretized PLRV typically obtained for a general mechanism?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer Bdkd7/10 · confidence 2/52024-07-13

Summary

This paper introduces a new method for improving the utility of privacy-preserving machine learning without sacrificing privacy protection. The authors develop efficient algorithms to calculate the trade-off curve between attack FPR and FNR using f-differential privacy (f-DP). They then show how to use this information to fine-tune the amount of noise added, allowing for precise control over privacy risks. This approach offers a more refined way to balance data usefulness and privacy in machine learning, addressing a key challenge in the field.

Strengths

The authors show that their direct calibration methods can significantly reduce the required noise scale compared to the standard approach, leading to an increase in utility (e.g., 18 percentage points higher accuracy) for the same level of privacy risk. They also demonstrate that calibrating for attack advantage (attack accuracy) can increase attack power in the low FPR regime, and show that calibrating for a desired FPR and FNR level mitigates this issue. By this method, the noise of DPSGD can be reduced tighter once we are aware of the privacy risk.

Weaknesses

The notation in this paper is heavy, could you provide notation tables? Could you provide a more detailed algorithm in the appendix about how to use $\epsilon$ , $\delta$, and q to generate Advantage Calibration’s Xω , Yω. Its quite hard to follow in Section 3’s demonstration. I think authors need an algorithm to demonstrate how they did the experiment from the supplement's code. I am willing to increase my score if the above questions can be answered.

Questions

see in weaknesses

Rating

7

Confidence

2

Soundness

4

Presentation

2

Contribution

4

Limitations

see in weaknesses

Reviewer JyJW5/10 · confidence 3/52024-07-28

Summary

Differential privacy (DP) mitigates privacy risks in machine learning by adding noise during training, balancing privacy and utility. Traditionally, the noise scale is set using a privacy budget parameter ε, which is then translated to attack risk. This two-step method often results in conservative risk assessments and reduced utility. The proposed approach directly calibrates noise to a desired attack risk, bypassing ε, thus decreasing the noise scale and enhancing utility while maintaining privacy. Empirical evidence shows that this method improves model accuracy for the same privacy level, offering a practical way to enhance privacy-preserving machine learning.

Strengths

1. The proposed framework successfully addresses technical challenges, providing compelling insights. 2. The experimental results are robust and strongly support the framework's effectiveness.

Weaknesses

1. The paper includes numerous definitions and symbols, which can be confusing for readers. Creating a table to summarize these terms and explain their meanings would greatly enhance clarity and help readers follow along more easily. 2. In Section 2, the problem statement is not clearly articulated, making it difficult to discern the main goal and the specific problem being addressed. Highlighting the main goal and explicitly defining the problem would make this section more readable and comprehensible.

Questions

1. In the experiment section, how about the results of other low FPR regimes?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

1. In Section 3.1, please include inference for ensuring advantage calibration in the appendix. 2. The two figures at the top of page 7 lack figure names. It appears they should be labeled as Figure 2.

Authorsrebuttal2024-08-12

As the author-reviewer period is coming to end, please let us know if we have addressed your concerns and if we can clarify anything else.

Reviewer pcXQ2024-08-13

Thanks for the rebuttal! I don't have further questions.

Authorsrebuttal2024-08-14

We would like to sincerely thank all reviewers for their constructive and actionable feedback, and for responsiveness during the discussion period. We believe the changes following the feedback, which we detailed in the individual responses and in the [general response](https://openreview.net/forum?id=hOcsUrOY0D&noteId=4H6Ec2IhLz), will significantly strengthen our paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC