acceptodds
Under review as a conference paper at ICLR 2027

Optimal Guarantees for Variational Auditing of Rényi Differential Privacy

17 views

Abstract

We study black-box auditing for machine learning algorithms that claim R\'enyi differential privacy (RDP) guarantees. We introduce an auditing framework, based on hypothesis testing, that directly estimates R\'enyi divergence between neighboring executions using the Donsker–Varadhan (DV) variational estimator. Our analysis yields explicit and non-asymptotic confidence intervals for RDP auditing via class-restricted DV estimators, separating statistical estimation error from algorithmic privacy leakage. We prove matching minimax lower bounds showing that, up to logarithmic factors, our sample-complexity guarantees are information-theoretically optimal, thereby establishing the first optimal guarantees for variational auditing of RDP. Empirically, we instantiate our framework for auditing DP-SGD in a fully black-box setting. Across MNIST and CIFAR-10, and over a wide range of privacy regimes, our auditors produce a strong overall improvement on empirical RDP lower bounds compared to prior state-of-the-art black-box methods especially at small and moderate R\'enyi orders where accurate auditing is most challenging.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Rényi Differential Privacy, Auditing Private Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BRthOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies black-box auditing of differentially private machine learning methods under Rényi DP. Using the variational representation of the Rényi divergence, and combine with the existing method on black-box auditing.

Critical strengths

This paper proves the sample complexity and demonstrate tighter results than the prior work.

Critical weaknesses

I think this paper missed an important literature [1]

  • The paper said we are the first to directly audit RDP, what about [1]? They also audit RDP.
  • This paper uses variational representation approach, [1] also use it.

[1] DP-Auditorium: a Large Scale Library for Auditing Differential Privacy. S&P 2024.

Questions for authors
  1. Can you clarify the technical difference between yours and DP-Auditorium?
  2. Does your algorithm need a single run or many run? In Alg 3, each run needs T iterations.
  3. I see your method has tigher estimation in Table 1-2. But does your method introduce extra overhead than the method you compare?
2 comments

We thank Reviewer NRth for their review and pointing us to relevant work. We are glad you found strength in our theoretical results. Please find our responses to your questions and concerns.


W1: DP-auditorium [1] also auditing Rényi-DP

Thank you for pointing us to this relevant work. We acknowledge the work of [1], and we cite it in section 1.1 of our paper. We mistakenly claim to be the “first” in the open review TL;DR, and in our contributions in which we have updated our manuscript. We again apologize for this mistake and hope it can be overlooked in consideration of our manuscript.


W2: Variational representation of the Renyi divergence used in DP-auditorium [1]

Thank you for this comment. The variational representation of the KL divergence originated in Donsker Varadhan’s representation of the KL divergence [2] and extended to the Rényi Divergence in [3]. Neural network based approaches are used in [4],[5]. These formulations for lower bound estimates are well known to the statistics community [1, section 1.1].


Q1: Technical differences between our work and DP-auditorium [1]

Thank you for asking for this clarification. Both works use the variational representation of the Rényi divergence to estimate RDP. Please find the main technical differences below from our paper below.

Theoretical Differences:

  • Our confidence bound holding over thee whole critic class vs theirs only holding for a singular critic: Our confidence bound holds over the whole critic class, while theirs only holds for a single fixed critic. Since the critic is chosen by looking at the data, their bound only applies if the critic is trained on one half of the samples and evaluated on the other half. Ours applies directly to the critic the trained neural network finds, using all the samples. This matters in realistic auditing settings (private learning such as dp-sgd), since it would mean setting aside half of the DP-SGD training runs to perform the audit which is very computationally expensive.

  • Matching lower bound: we also prove a lower bound in theorem 3 showing that any confidence bound that holds over the whole critic class needs at least Ω(d/ε2)\Omega(d/ \varepsilon^2) samples, so the rate in theorem 2 cannot be improved. [1] does not provide a lower bound and we cannot tell if their sample complexity is tight or not. We also provide more general upper and lower bounds in the appendix for our auditor in Theorems 1,5, and 6 in our manuscript.

  • Audit algorithms output: Our auditing algorithm outputs a lower bound ε\varepsilon estimate, while [1] outputs a pass / fail against a certain threshold. A pass / fail only tells us if the divergence is above the tested threshold, meanwhile a lower bound ε\varepsilon estimate gives us more insight to how much privacy a mechanism leaks.

Experimental Differences

  • Private mechanisms audited: We audit DP-SGD on CNNs trained with MNIST and CIFAR-10. This is the algorithm whose privacy we are the most interested in and is the gold standard algorithm to perform DP audits on, see [6][7][8][9][10][11][12][13][14]. [1] audits very simple private mechanisms: Laplace and Gaussian means, sparse vector technique, noisy max, and one flawed simple private gradient descent implementation.
  • Regime of samples: for their experiments [1] considers 500,000 samples to perform an audit, while we only use 500. We chose 500 as it provides a tradeoff between computational constraints for private mechanisms utilized in realistic scenarios and achieving tight audits.
  • Baseline comparisons: We compare to previous state of the art black box auditing methods, showing our methods produce tighter black box audits and overall outperform the current state of the art [8]. DP-auditorium does not consider other external state of the art baselines when performing their audits.

Q2: Algorithm runs: one or multiple?

Thank you for the question. Our auditor operates in the black-box (where we can only observe the final outputs) regime where producing a tighter audit requires several iterations. This is standard in black box auditing when attempting to produce a tight lower bound audit, see [7][8], and see [9][10] where only one run is done but a loose audit is produced.


Q3: Extra overhead when compared to the current SOTA method.

Both methods use the same number of blackbox samples, the main overhead cost in performing the audit (the T iterations). The operations after introduce negligible overhead (over 50 GPU hours on an A100 cluster vs seconds for everything else for both audits) We train a small NN after to estimate the Rényi divergence, while [8] uses a decision threshold to obtain FPR/FNR results which are converted to confidence bounds which are used to calculate a lower bound GDP estimate which is converted to RDP.


Thank you again for your feedback as well as time taken to improve our manuscript. We hope that we have resolved your concerns regarding differences between DP-auditorium [1] and would consider raising your score if so. Please let us know if you have any additional questions, we are happy to answer while the rebuttal period is still open.


References

[1] Kong et al. DP-Auditorium: a Large Scale Library for Auditing Differential Privacy. S&P 2024.

[2] Donsker and Varadhan. Asymptotic evaluation of certain Markov process expectations for large time, IV. Communications on Pure and Applied Mathematics, 1983.

[3] Anantharam. "A variational characterization of Rényi divergences." IEEE TIT

[4] Belghazi et al. "Mutual information neural estimation." ICML 2018.

[5] Birrell et al. "Variational representations and neural network estimation of Rényi divergences." SIMODS 2021.

[6] Jagielski et al. "Auditing differentially private machine learning: How private is private sgd?" NeurIPS 2020.

[7] Nasr et al. "Tight auditing of differentially private machine learning." USENIX Security 2023. Distinguished Paper Award.

[8] Muthu Selva Annamalai et al. "Nearly tight black-box auditing of differentially private machine learning." NeurIPS 2024.

[9] Steinke et al. Privacy Auditing with One (1) Training run. NeurIPS 2023. Best Paper Award.

[10] Mahloujifar et al. "Auditing f-differential privacy in one run." ICML 2025.

[11] Cebere et al. Tighter Privacy Auditing of DP-SGD in the Hidden State Threat Model. ICLR 2025.

[12] Pillutla et al. Unleashing the Power of Randomization in Auditing Differentially Private ML. NeurIPS 2023.

[13] Feng and Tramèr. Privacy Backdoors: Stealing Data with Corrupted Pretrained Models. ICML 2024.

[14] Xiang et al. Privacy Audit as Bits Transmission: (Im)possibilities for Audit by One Run. USENIX Security 2025.

Reviewer_ZhG4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a black-box auditing framework for Rényi Differential Privacy (RDP). It uses a neural-network-based variational estimator to estimate Rényi divergence between the outputs of a learning algorithm on neighboring datasets. The authors provide finite-sample confidence bounds and minimax lower bounds for the estimator. Experiments on DP-SGD using MNIST and CIFAR-10 show improvements over a previous black-box auditing method, especially at small Rényi orders.

Critical strengths

Directly auditing RDP is valuable, since RDP is widely used for privacy accounting. The variational estimation approach is interesting and could provide an alternative to traditional attack-based auditing.

Critical weaknesses

1. A serious concern about the theoretical guarantees.

Corollary 3 claims that a lower bound on Rényi divergence can be converted into a certificate of violation of (ϵ,δ)(\epsilon,\delta)-DP. However, this does not seem correct. RDP implies an approximate DP guarantee, but the reverse implication does not generally hold. A mechanism can satisfy approximate DP while having arbitrarily large Rényi divergence. Therefore, violating an RDP bound does not necessarily imply violating the corresponding approximate DP guarantee. I am also not convinced by the claimed optimality result. Theorem 3 establishes a lower bound for uniformly estimating the entire variational objective over the critic class, whereas auditing only requires estimating its supremum. These are different problems. In fact, the constructed distribution family appears to have the same supremum value across hypotheses. It is unclear how this proves optimal sample complexity for RDP auditing itself.

2. The experiments do not demonstrate the claimed confidence guarantees.

One of the paper's main contributions is providing statistically valid confidence bounds. However, the experiments mainly report empirical divergence estimates with standard deviations. They do not evaluate the actual confidence intervals derived in the theory. Are these bounds meaningful with only 500 observations and a neural-network critic? The theoretical error depends on the critic dimension and other constants, so the resulting bounds could be quite loose. Some results also exceed the nominal theoretical privacy parameters. For example, Table 8 reports an estimated RDP of 1.512 when the target is 1.0. This does not necessarily indicate a privacy violation, but raises questions about estimation bias and statistical reliability. The authors should evaluate the coverage and tightness of their confidence intervals, ideally on simple mechanisms where the true Rényi divergence is known.

3. The experimental evaluation and baseline comparisons are limited.

The experiments only evaluate DP-SGD on MNIST and CIFAR-10 using CNNs, with Rényi orders limited to 1.25 and 2. It is unclear how well the method works at larger orders or on other privacy mechanisms. The comparison is also limited. Since the observed outputs are scalar losses, it would be useful to compare against simpler divergence estimators, such as histogram-based or density-estimation methods. Furthermore, the improvements are not consistent across all settings. For example, at α=2\alpha=2 and target RDP 10, the previous method substantially outperforms the proposed method. These limitations should be discussed more carefully.

Questions for authors
  1. Is the DP violation certificate in Corollary 3 correct? How does a lower bound on RDP imply a violation of (ϵ,δ)(\epsilon,\delta)-DP? Please clarify the logical implication and the normalization of the Rényi divergence.

  2. What exactly is optimal about the proposed estimator? Theorem 3 concerns uniform estimation of the variational objective. Does the same lower bound apply to estimating the scalar Rényi divergence or its class-restricted supremum?

  3. Can the authors evaluate the actual confidence bounds? What are the numerical confidence intervals for the reported experiments? Are they non-vacuous with 500 samples? Can the authors demonstrate empirical coverage on mechanisms with known Rényi divergence?

  4. Can the authors expand the experiments? How does the method perform at larger Rényi orders, with alternative divergence estimators, or on privacy mechanisms other than DP-SGD? Why does its performance become worse than the baseline in some settings?

3 comments

We thank Reviewer ZhG4 for their feedback and review. We are glad you found our approach interesting and novel. Please find our responses to concerts and questions.

W1/Q1 Corollary 3:

The reviewer is right. Our proof reverses the implication from RDP to approximate DP. An upper bound on Renyi divergence gives an approximate DP guarantee but exceeding that RDP bound does not imply a violation of the corresponding approximate DP guarantee. We have removed the approximate DP conclusion from Corollary 3 and the corresponding position of its proof.

However, a valid consequence is a certificate of an RDP violation. Specifically, using the normalization Dα=αRαD_\alpha=\alpha R_\alpha (introduced in Section 2.1), let rn(β)r_n(\beta) be a valid uniform confidence radius for the empirical variational objective. Then Dα(Q∣P)≥α(R^α,nΘ(Q∣P)−rn(β))D_\alpha(Q|P) \ge \alpha\left(\widehat R_{\alpha,n}^{\Theta}(Q|P)-r_n(\beta)\right) with probability at least 1−β1-\beta. Consequently, rejecting a (α,ρ)(\alpha,\rho)-RDP guarantee when this lower confidence bound exceeds ρ\rho has Type-I error at most β\beta. This follows from concentration and the variational lower-bound (without any conversion to approximate DP). All of our experiments reflect RDP levels and we have updated the paper to reflect these changes.


W1/Q2 Theorem 3:

We also agree that uniformly estimating the function θ↦V(θ)\theta\mapsto V(\theta) and estimating its supremum are different statistical tasks. Our proof implies a lower bound for the former. As the reviewer has observed, in the constructed family the maximizing critic depends on the hidden hypothesis, whereas the supremum value is identical across hypotheses. We have retained our result as a lower bound for uniform estimation of the variational objective, explicitly stated this limitation, and removed any claims that it implies optimal sample complexity for RDP auditing. We have also updated the title, abstract, contributions, Section 3.3, and Remark 1 accordingly.


W2/Q3: Numerical confidence intervals meaningful and empirical coverage

Thank you for your feedback regarding our experimental section. Theorem 2 gives a uniform bound over the whole critic class with a dependence on dd that theorem 3 shows is necessary. When using realistic critic size, privacy levels, and computational cost parameters the confidence radius rn(β)r_n (\beta) from theorem 2 exceeds the εα\varepsilon_\alpha values in the tables for this paper. We emphasize that the audit would remain valid but more conservative with applying the confidence intervals.

Moreover, Table 8 also reports an SD of 0.662, showing that our audit remains within the standard deviation. This also occurs in our baseline [3] (section 5.3) sometimes. In our case, we attribute this error to the estimation error and nonexactness from using neural networks to estimate the Rényi divergence, and the different gradient samples from training on a smaller dataset given to our estimator.

Lastly, see below for an ablation where the true Rényi divergence is known, showing our audit lower bound reamians valid but more conservative in cases.


W3/Q4: Limited experimental evaluation

Thank you for your comment. When auditing differential privacy, the private mechanism one is most interested in auditing is DP-SGD. It is standard to audit this algorithm in papers, instead of simpler mechanisms, see [1][2][3][4][5][6][7][8][9]. We agree it would be interesting to expand our experiments to private synthetic data auditing since our estimator allows inputs of different sizes, but that is left for future work and is out of the rebuttals scope.

To explain our non-consistent improvement over the previous SOTA, see TyZj W2.

Alternate divergence estimators are our of the scope of this work, as we focus on Rényi-DP. We demonstrate coverage where the true Rényi divergence is known. We report the estimated Rényi divergence between 10 dimensional Gaussians N(0,Id)N(0, I_d) and N(Δe1,Id)N(\Delta e_1, I_d) as a function of α∈(2,5)\alpha \in (2, 5) for values of Δ∈0.5,2,5\Delta \in {0.5, 2, 5}, and compare it to the exact value αΔ2/2\alpha\Delta^2/2. As one can see from the results, the lower bound audit becomes more conservative but remains valid. This is due to the non exactness when estimating these information-theoretic divergences with neural networks.

Results for 10-D Gaussians

α\alphaΔ\DeltaTrue DαD_\alphaDα^\widehat{D_\alpha}
2.00.50.250.224
2.02.04.03.256
2.05.02520.844
5.00.50.6250.486
5.02.0106.515
5.05.062.543.245

Again we sincerely thank you for your time and feedback. We appreciate your concern towards our theoretical guarantees helping us improve that portion of the manuscript. We hope we have addressed all your questions and issues, and would consider raising your score if so. Please let us know if any additional comments or concerns arise.

References

[1] Jagielski et al. "Auditing differentially private machine learning: How private is private sgd?" NeurIPS 2020.

[2] Nasr et al. "Tight auditing of differentially private machine learning." USENIX Security 2023. Distinguished Paper.

[3] Muthu Selva Annamalai et al. "Nearly tight black-box auditing of differentially private machine learning." NeurIPS 2024.

[4] Steinke et al. Privacy Auditing with One (1) Training run. NeurIPS 2023. Best Paper.

[5] Mahloujifar et al. "Auditing f-differential privacy in one run." ICML 2025.

[6] Cebere et al. Tighter Privacy Auditing of DP-SGD in the Hidden State Threat Model. ICLR 2025.

[7] Pillutla et al. Unleashing the Power of Randomization in Auditing Differentially Private ML. NeurIPS 2023.

[8] Feng and Tramèr. Privacy Backdoors: Stealing Data with Corrupted Pretrained Models. ICML 2024.

[9] Xiang et al. Privacy Audit as Bits Transmission: (Im)possibilities for Audit by One Run. USENIX Security 2025.

Thank you for the detailed response and clarifications. I appreciate the authors' efforts in addressing my concerns, especially acknowledging and correcting the issues with Corollary 3 and the sample complexity claim in Theorem 3.

However, I still have concerns about the paper's main contributions. In particular, the authors acknowledge that the theoretical confidence bounds are too loose to be meaningful in their experimental settings. Since finite-sample confidence guarantees are a central contribution of this paper, I think this is a significant limitation.

Overall, while the rebuttal clarifies several issues, I do not think it sufficiently addresses my main concerns about the theoretical and empirical contributions. I will therefore maintain my original rating and leave it to the AC to weigh these remaining concerns against the paper's contributions in the final decision.

Reviewer_TyZjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper shows a techinque for auditing Renyi Differential Privacy (RDP) in the black-box setting, where the attacker does not have access to the model's gradients. The lower bound on RDP is obtained via a Donsker-Varadhan (DV) variational estimator of of Renyi-Divergence. The estimate is obtained from multiple observations of model loss on neighboring datasets. The authors provide finite-sample upper and lower confidence bounds on the estimator. Empirically, they show that their method can obtain tigher lower bounds compared to prior work on CIFAR-10 and MNIST datasets.

Critical strengths
  • The paper directly targets RDP auditing, as opposed to the standard approximate DP.
  • The paper shows conistency and finite sample confidence interval guarantees for the estimator.
  • Empirical results show improvement over state of the art, which is especially hard to achieve in the black box setting (although improvements do not hold in all experimental settings).
Critical weaknesses
  • The estimator can become unstable for privacy values alpha > 2. The authors run their experiments for alpha in (1, 2].
  • The improvement over prior state of the art is not consistent in all experimental settings (the authors experiment with different values of clipping norm and number of private samples).

Minor

  • In the experiments in the appendix, I would recommend specifying the clipping norm and private dataset size for each table. Otherwise the experimental setup is not fully clear.
Questions for authors
  • In Theorem 2, the confidence interval holds for the estimator sup⁡θ∈ΘV^n(θ)\sup_{\theta \in \Theta} \hat{V}_n(\theta), however the estimator in algorithm 1 is obtained from some averaging of V^n(θ)\hat{V}_n(\theta). How should one think about this gap between Algorithm 1 and Theorem 2?

  • In Lines 409-411 it says that you non-privately fine-tune the model with the remaining half of the dataset. Did you mean "privately" fine-tune? I was confused by these lines.

  • In Line 414, you say that you set the batch size equal to dataset size for the private fine-tuning. This seems like an extreme choice. Was it to obtain better accuracy for the model?

2 comments

We first thank reviewer TyZj for their comments and feedback. We are happy you found strength in our theoretical finite sample guarantees and empirical results. Kindly find our response to your questions and concerns below.


W1: Experimental results for α∈(1,2]\alpha \in (1,2].

We thank you for bringing this up. We acknowledge that instability may exist with a variational Rényi estimator since estimating these information-theoretic divergences with neural networks may lead to non-exactness as seen in [1, 2]. In similar literature (https://arxiv.org/pdf/2007.03814 ([2] fig. 2), https://arxiv.org/pdf/1801.04062 ([1] fig. 1)), parameters are limited to smaller dimensions to provide tighter estimates, with smaller alpha values due to the disconnect between theory and practice of neural estimators.

To validate that our estimator holds up in practice in producing a lower bound (which may be more conservative) at higher α\alpha values, we report the estimated Rényi divergence between 10 dimensional Gaussians N(0,Id)N(0, I_d) and N(Δe1,Id)N(\Delta e_1, I_d) as a function of α∈(2,5)\alpha \in (2, 5) for values of Δ∈0.5,2,5\Delta \in {0.5, 2, 5}, and compare it to the exact value αΔ2/2\alpha\Delta^2/2. We run this experiment with higher dimensional variable estimates with neural networks which may lead to less tight results, due to non-exactness. In our paper’s auditor setting we expect less of this, since the variables are 1-D rather than 10-D.

Results for 10-D Gaussians

α\alphaΔ\DeltaTrue DαD_\alphaDα^\widehat{D_\alpha}
2.00.50.250.224
2.02.04.03.256
2.05.02520.844
5.00.50.6250.486
5.02.0106.515
5.05.062.543.245

W2: Inconsistent improvement over SOTA:

Thank you for your comment. The inconsistent (but overall) improvement can be attributed to two main factors. One is again the non-exactness for estimating the variational Rényi divergence via neural networks. Most of the time we fail to beat the SOTA baseline at higher ε\varepsilon values. This is the second factor and stems from the exponential nature of the Rényi divergence. Referencing Birrell’s equation in [2], this is the log of the exponential moment of the privacy loss. Any direct Rényi divergence estimator must average a heavy-tailed variable whose mean is dominated by exponentially rare samples, beyond our 500500 total samples at εα=10\varepsilon_\alpha = 10. Since the tailed event mean variables are currently out of scope this biases our empirical estimate downward, making the audit more conservative but still valid. Given this, increasing εα\varepsilon_\alpha would widen the SOTA’s advantage in this regime. To remedy this our estimator would need more samples, but we decided on 500 for our paper to create a realistic scenario.

Moreover, our auditor performs well in the practical regime. See [3], where recommendations for selecting ε\varepsilon in practice are provided (ε≤10\varepsilon \le 10 for non-weak/informal guarantees). Our experimental epsilons also satisfy [0.5,1,2,2,10]μ[0.5, 1, \sqrt{2}, 2, \sqrt{10}] \mu-GDP and (ε,10−5\varepsilon, 10^{-5})-DP, where ε\varepsilon is [2, 4.38, 6.57, 10, 17.85], showing that we achieve SOTA in the practical training regime of ε≤10\varepsilon \le 10.


Q1: Gap between Algorithm 1 and Theorem 2 (V^n(θ)\widehat V_n(\theta) and sup⁡θ∈ΘV^n(θ)\sup_{\theta \in \Theta} \widehat V_n(\theta))

Thank you for this question. There are two things bridging the gap, the exponential moving average (EMA) and optimization error. For the EMA/averaging, we use 0.99 of the new value/gradient estimate and 0.01 of the old value/gradient, making the averaging quantity negligible. Furthermore, any optimization error, bridging V^n(θ)\widehat V_n(\theta) and sup⁡θ∈ΘV^n(θ)\sup_{\theta \in \Theta} \widehat V_n(\theta) only makes the lower bound audit more conservative (and still valid) since the supremum is not reached.


Q2: lines 409-411 non-private fine tuning for worst case initialization

Thank you for pointing this out. In this context “non-privately” is correct. To perform our worst-case initialization we split the dataset into two halves. One for non-private pre-training and the other for private training, following the approach in [4]. This approach of pre-training (and non-privately fine tuning done here) produces initial parameters under which the ordinary samples already have small gradients, so when private training begins the canary’s gradient dominates, which yields the tightest audits.


Q3: Large batch size in experiments

Thank you for your question. Although in some settings it may seem extreme, setting the batch size to the dataset size in black-box auditing is done to achieve tighter audits, see similar done in [4][5] and [6][7] highlight advantages of large batch sizes.


We will also specify the clipping norm and private dataset size for each table. Thank you again for your comments and feedback as well as time taken to improve our manuscript. We hope that we have resolved all your concerns and would consider raising your score if so. Please let us know if you have any additional questions, we would be happy to answer.


References

[1] Belghazi et al. "Mutual information neural estimation." ICML 2018.

[2] Birrell et al. "Variational representations and neural network estimation of Rényi divergences." SIMODS 2021.

[3] Ponomareva et al. “How to DP-fy ML: A Practical Guide to Machine Learning with Differential Privacy.” JAIR 2023.

[4] Muthu Selva Annamalai et al. "Nearly tight black-box auditing of differentially private machine learning." NeurIPS 2024.

[5] Nasr et al. "Tight auditing of differentially private machine learning." USENIX Security 2023.

[6] Dörmann et al. "Not all noise is accounted equally: How differentially private learning benefits from large sampling rates."

[7] De et al. "Unlocking high-accuracy differentially private image classification through scale."

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.