DP-Mix: Mixup-based Data Augmentation for Differentially Private Learning

Data augmentation techniques, such as simple image transformations and combinations, are highly effective at improving the generalization of computer vision models, especially when training data is limited. However, such techniques are fundamentally incompatible with differentially private learning approaches, due to the latter's built-in assumption that each training image's contribution to the learned model is bounded. In this paper, we investigate why naive applications of multi-sample data augmentation techniques, such as mixup, fail to achieve good performance and propose two novel data augmentation techniques specifically designed for the constraints of differentially private learning. Our first technique, DP-Mix_Self, achieves SoTA classification performance across a range of datasets and settings by performing mixup on self-augmented data. Our second technique, DP-Mix_Diff, further improves performance by incorporating synthetic data from a pre-trained diffusion model into the mixup process. We open-source the code at https://github.com/wenxuan-Bao/DP-Mix.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer DhpR5/10 · confidence 3/52023-06-27

Summary

The paper proposes two data augmentation techniques, MIXUP-SELFAUG and MIXUP-DIFFUSION, for differentially private learning. The authors investigate why naive applications of multi-sample data augmentation techniques, such as mixup, fail to achieve good performance and propose these two techniques specifically designed for the constraints of differentially private learning. MIXUP-SELFAUG performs mixup on self-augmented data, which results in lower gradients norm average and variance leading to smoother training. MIXUP-DIFFUSION further improves performance by incorporating synthetic data from a pre-trained diffusion model into the mixup process. The paper also discusses the challenges of applying data augmentation techniques in the context of differential privacy and compares the proposed techniques with existing methods on various datasets. The experimental results show that the proposed techniques achieve state-of-the-art classification performance across a range of datasets and settings. Overall, the paper proposed effective data augmentation techniques that can improve the performance of machine learning models while preserving the privacy of sensitive data.

Strengths

This paper is easy to read. This paper considerately used DP-SGD for differentially private learning. These techniques improve the performance of mixup-based data augmentation. The paper also discusses the challenges of applying data augmentation techniques in the context of differential privacy.

Weaknesses

1. The absence of a theoretical analysis for MIXUP-SELFAUG in the submitted work is notable. The authors have provided an intuition on why mixup could harm differential privacy and presented an alternative approach to mitigate this issue. It could further strengthen the paper if the authors could provide a theoretical guarantee for differential privacy. 2. In relation to MIXUP-SELFAUG which doesn't mix labels, questions arise about why MIXUP-SELFAUG improves performance. The original intent of mixup, which is label interpolation for better performance, seems not fully addressed in this context. The authors could consider explaining their approach's effectiveness via intuition or mathematical formulation. [1] Zhang, Linjun, et al. "How does mixup help with robustness and generalization?." ICLR 2021 [2] Jeong, Jongheon, et al. "Smoothmix: Training confidence-calibrated smoothed classifiers for certified robustness." NeurIPS 2021 [3] Zhang, Linjun, et al. "When and how mixup improves calibration." ICML 2022 [4] Park, Chanwoo, Sangdoo Yun, and Sanghyuk Chun. "A unified analysis of mixed sample data augmentation: A loss function perspective." NeurIPS 2022 [5] Zou, Difan, et al. "The benefits of mixup for feature learning." arXiv preprint arXiv:2303.08433 (2023). [6] Oh, Junsoo, and Chulee Yun. "Provable Benefit of Mixup for Finding Optimal Decision Boundaries." ICML 2023 3. Concerning the implementation of the diffusion model, there are a few considerations that the authors might need to address. The necessity for incorporating a diffusion model is not sufficiently justified in the current draft. For instance, the diffusion model used in this study is pretrained, which could potentially introduce the element of knowledge distillation. This raises the question: Is the observed performance improvement mainly due to distillation effects? Additionally, the use of a diffusion model not trained with DP-SGD could potentially pose issues for differential privacy learning. The authors are encouraged to provide clarification on these points. 4. While the adoption of a diffusion model might benefit differential privacy, the paper could delve deeper into explaining why it's advantageous for ensuring differential privacy. A detailed discussion on this matter could provide readers with a better understanding of the authors' choice and its implications on the overall study. The authors' insights on this matter would be greatly appreciated. [1] Chourasia, Rishav, Jiayuan Ye, and Reza Shokri. "Differential privacy dynamics of langevin diffusion and noisy gradient descent." NeurIPS 2021 [2] Dockhorn, Tim, et al. "Differentially private diffusion models." arXiv preprint arXiv:2210.09929 (2022). [3] Ghalebikesabi, Sahra, et al. "Differentially Private Diffusion Models Generate Useful Synthetic Images." arXiv preprint arXiv:2302.13861 (2023). [4] Lyu, Saiyue, et al. "Differentially Private Latent Diffusion Models." arXiv preprint arXiv:2305.15759 (2023).

Questions

I want to clarify especially the second and third issues of the weakness. If this is well-addressed, I am down to re-evaluate this paper.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors adequately addressed the limitation and potential negative societal impact on their work.

Reviewer DhpR2023-08-10

Thank you.

Thank you for your detailed explanation. Regarding the first point, I appreciate your clarification. However, I am not sure that the sensitivity difference between 2C and C might not be significant given that it's a constant factor. Plus, it is not a worst-case guarantee, so it would be more clarified. For the second point, I might need further clarity. In the third link you shared, they utilized a "contrastive" loss to find a good representation. However, it appears this paper does not employ the same approach. I'm curious to understand how this work might benefit from the same label without using contrastive loss. As for points 3, 4, and 5, my concerns remain. My understanding is that the diffusion model isn't a "differential private" method since it hasn't been trained with DP-SGD. Could you explain this? If a model isn't trained using a "differential private" method, wouldn't it naturally exhibit improved performance? I agree with your point on the alignment of training data and the testing with a synthetic dataset. I'm not sure that the absence of data leaks isn't necessarily the same as ensuring differential privacy. (Using the same dataset also might induce not good differential privacy. Not really related I believe) Thank you once again for your time and dedication in addressing my queries.

Authorsrebuttal2023-08-11

Thank you for your quick response. We would like to clarify some points: **1.** The difference between C and 2C is massive because it implies effectively doubling the noise level to achieve the same privacy guarantee with DP-SGD, which results in much worse performance. For example, see Table 2 of our submission PDF for EuroSAT, where for $\varepsilon=1$ and $\varepsilon=2$ correspond to almost doubling the noise level (to $\sigma=25$ from $\sigma=13$) with all other parameters (batch size, number of epochs, etc.) remaining the same (except the learning rate which is tuned to give best results in each case). As can be seen in Table 2, the absolute test accuracy decrease resulting from this almost doubling of the noise level is over $5$%. Also, to clarify: our analysis for the doubling of sensitivity is a worst-case analysis. This is because differential privacy is a worst-case notion, as it applies to all pairs of neighboring datasets. **2.** Here we want to clarify two separate points: - **a.** Most data augmentation techniques, such as flipping and cropping, operate on a single example and thus do not mix labels, yet they still improve generalization. Mixup-SelfAug operates in a similar fashion. - **b.** The related works mentioned in our previous response are not meant to endorse any particular approach or suggest that we would adopt it. Our point is simply that empirical results (from these related works) show that mixup without label mixing still enhances generalization. **3.** Thank you for clarifying. We think we understand what you are asking now. If we train a model with DP-SGD, the differential privacy protection is always with respect to the training set used. This means that if we take a model already pre-trained on dataset X and then fine-tune it with DP-SGD on dataset Y, the differential privacy guarantee only applies to dataset Y; there is no guarantee on dataset X unless pre-training was also done using DP-SGD. This is why in our paper we consider both training from scratch and fine-tuning. In the training from scratch setting, we consider only Mixup-SelfAug and not Mixup-Diffusion. In the fine-tuning/pre-trained setting, we assume that the dataset used to pre-train the model (LAION-2B) is public so DP-SGD pre-training is not necessary. This is a standard assumption from prior works. For example, De et al. [10] pretrain their models on ImageNet, JFT-300M, and JFT-4B. In fact, the paper that introduced DP-SGD, Abadi et al. [2], pretrain their model on CIFAR-100. In our case, we specifically pretrain ViT and ConvNext on the LAION-2B dataset so that both the model to be fine-tuned and the diffusion model used the exact same training data. Given the assumption that this dataset is public, we do not need to do the pretraining with DP-SGD. When we fine-tune the pretrained model with Mixup-SelfAug or Mixup-Diffusion differential privacy protects the fine-tuning data.

Reviewer DhpR2023-08-11

Thank you

For the first point, I understand that there might be a significant difference between C and 2C. Thank you for clarifying that. Regarding the second point, I don't necessarily agree. The purpose of mixup is to blend both labels and data. If we don't blend the labels, there would be no reason to use mixup; instead, we'd just employ traditional augmentation techniques like flipping. Mixup is done for smoothing the trained function so label mixing is necessary to have these effects. And smoothing trained function actually provides good generalization. (See Figure 1 in https://arxiv.org/pdf/1710.09412.pdf) My contention is that the effects observed might not stem from the "mixup" method per se, but could be attributed to having a large dataset or incorporating translations and rotations. How can we clarify this distinction? I've fully grasped the third point in the sense that previous works also use pre-trained models, while do not fully agree that this is a valid approach. But, For point 3, I think this is enough to discuss (it is a really fundamental question to the literature). Thank you for the detailed explanation.

Authorsrebuttal2023-08-16

**Is mixup without label mixing similar to traditional augmentation?** We agree with the reviewer that it is important to consider whether mixup without mixing labels could have a similar effect as other traditional augmentations such as translations and rotations. Therefore, we performed additional experiments using several widely-used augmentation techniques ([https://arxiv.org/abs/2002.05709](https://arxiv.org/abs/2002.05709) -- Colorjitter, Translations and Rotations, Cutout, Gaussian Noise, Gaussian Blur, and Sobel Filter) in addition to Self-Aug to compare them to Mixup-SelfAug. The table below shows the results for each augmentation separately with the last row combining all augmentations together (randomly chosen). We observe that these augmentations do *not* enhance performance to the same degree as Mixup-SelfAug. In a few cases, the augmentations improve performance slightly above Self-Aug, but some of them also diminish performance. We are thankful for the suggestion and will add this table to the revised paper. **Table:** Vit-B-16 model performance on CIFAR-10 and CIFAR-100 with $\varepsilon$=1. | Method | CIFAR-10 | CIFAR-100 | |---------------------------------------|----------|-----------| | Self-Aug | 96.49% (±0.12%) | 79.28%(±0.18%) | | Mixup-SelfAug | 97.21%(±0.27%) | 81.75% (±0.15%) | | Self-Aug + Colorjitter | 96.30% (±0.09%) | 78.15% (±0.21%) | | Self-Aug + Translations and rotations | 96.17%(±0.14%) | 73.87% (±0.18%) | | Self-Aug + Cutout | 96.50% (±0.10%) | 79.55% (±0.11%) | | Self-Aug + Gaussian Noise | 95.84%(±0.21%) | 73.47%(±0.17%) | | Self-Aug + Gaussian Blur | 96.72%(±0.16%) | 79.95% (±0.13%) | | Self-Aug + Sober Filter | 72.47%(±0.11%) | 9.54% (±0.22%) | | Self-Aug + All (Random) | 96.45%(±0.15%) | 79.42% (±0.14%) |

Reviewer DhpR2023-08-16

Quick question

So these results are all about using the same number of samples? Thank you very much. I will adjust my score.

Authorsrebuttal2023-08-17

Yes, for all Self-Aug we use k=16 and for Mixup-SelfAug k=k'=16. Regarding the number difference between Self-Aug and Mixup-SelfAug, please check point 3 of our rebuttal and Table 6 of our attached one page pdf, it shows that increasing k cannot improve performance for Self-Aug.

Reviewer DhpR2023-08-17

Thank you

I changed my score again because now every point is clear. Even though this cannot make the theoretical explanation (about why mixing the same label helps), I think this holds empirically by various experiments that the authors provided.

Reviewer sSKD5/10 · confidence 5/52023-07-04

Summary

This paper considers the privacy-utility tradeoff for ML models trained with differential privacy guarantees, and develops a technique using data augmentation on image datasets to train models with high accuracies on standard benchmarks with DP guarantees. Mixup, a commonly used augmentation technique in computer vision, and its variants are not compatible with standard DP-SGD because a single datapoint could influence many training instances through augmentations. Recent work [10] pointed out that augmentations involving a single datapoint can be used if clipping if done after averaging all gradients stemming from augmentations of the datapoint. The present paper develops two methods inspired by mixup that involve a single datapoint and therefore can be used with DP-SGD. Experiments show performance comparable to state-of-the-art methods for training DP models.

Strengths

Thank you to the authors for sharing their interesting research! Overall, the paper tackles an important subject - improving the utility of models with rigorous privacy guarantees. This is an active area of research, and the paper engages with recent work appropriately. The proposed methods are simple to implement for image datasets and provide a boost in utility. An attempt is made to explain why mixup methods improve performance by checking the distributions of gradient norms. Originality: The paper develops two variants of mixup that can be applied to DP-SGD training. I do not know of prior work using mixup for DP-SGD. Quality: The analysis of differential privacy is careful and correct. The authors are diligent in highlighting parts of the analysis which are particularly tricky (L91, etc.) although these aspects are known in the literature. Relevant and recent methods are used for benchmarking, and error bars over 3 runs are given. Source code is provided. Clarity: The paper is quite clear. Significance: The poor utility of models trained with DP-SGD holds back the adoption of this important privacy enhancing technology, so new methods (like the ones in this paper) that improve the utility are necessary.

Weaknesses

Originality: The paper combines the augmentation techniques developed for DP in [10], with the idea of mixup which is very standard. The combination is only slightly non-trivial in this context. Quality: The experiments lack detail - it would not be possible to reproduce the author's work with the current state of the paper and appendices. Prior SotA work [10] is frequently referenced, but performance levels from that paper are not reproduced or surpassed. The experiment using diffusion models is not benchmarked fairly. Broader impacts are relegated to the appendix. Clarity: There are minor grammatical and typesetting issues, but they do not hurt the clarity of the paper. Figure 2 requires clarification. Significance: The applicability of this technique is limited to image datasets, whereas that of [10] for example can be applied much more broadly.

Questions

1. What augmentations were used in the experiments? I have not found this detail clearly stated in the paper or appendices, despite it being a central focus. L167 mentions a "randomized transformation function $T$" but this is not brought up again. L247 mentions flipping and cropping, but it seems to imply these are not the only augmentations used. The lack of important details such as this negatively impact the reproducibility of this research. 2. Can the authors state what computing resources were used and information on the runtime of their methods in comparison to prior work? 3. Can the authors give details on the diffusion model that was used (architecture, how it was trained, etc.)? All we know is that it was from Open Clip and trained on LAION-2B (L218). 4. The experiments attempt to follow [10] in many ways for close comparison, but the results reported in Section 5 as baselines do not reach the level of that paper. For instance, [10] reports 81.4% accuracy on CIFAR-10 training from scratch with a WRN-16-4 and epsilon=8. Table 1's Baseline is presumably meant to represent the method from [10], but only achieves 72.5% accuracy. The author's proposed method achieves 78.7% accuracy in the same setting. Given that the authors make claims of SotA performance, can they clarify why the results reported in [10] were not directly used or reproduced? 5. There are some results which are repeated across tables (e.g. first two rows of Table 4 are reproduced from Table 3). Could this be mentioned explicitly where applicable so that it is easier to track and compare results? A reader moving quickly might assume they are separate results. 6. In Table 4 the Mixup-Diffusion method uses additional data compared to the other approaches, namely the data generated by the diffusion model. Since the utility of DP models is often constrained by access to data, it is important to use baselines that have the same level of data access. Have the authors put thought into an appropriate baseline technique in the style of DP-SGD that also makes use of the synthetic data? (The most naive option being to pre-train or extend the training dataset with the same synthetic data made available to the mixup-diffusion method.) 7. In Figure 2 it is not clear what the difference is between the curves and the blue bars. I interpret each curve as the distribution of gradient norms over individual datapoints in the training dataset at a particular epoch (these details are not given, I have to guess). However, the blue bars are stated to be a histogram of the average norm of gradients, but what is the average taken over and what is the resulting distribution over? I note that the paper at various times considers epochs, minibatches, and microbatches, so the authors should clearly state which notion they are using in their analysis. 8. Also for Figure 2, the authors do not mention in the paper or appendices which dataset is used. Can the authors provide sufficient detail to reproduce their experiments (without having to dig through source code)? 9. The result in Figure 2 showing lower average gradient norms is indeed suggestive that mixup leads to smoother training and faster convergence. However, have the authors found an explanation for why mixup leads to lower average gradient norms? This is not at all self-evident, especially for Mixup-Diffusion where training and synthetic datapoints are mixed, and those images can come from different classes (as illustrated in Fig 1). 10. Have the authors checked the equivalent of Figure 2 for Mixup-Diffusion, or for other datasets? Can this analysis be included? Minor: In L67, couldn't $\delta=0$ be a valid choice, and for that matter $\delta=1$ corresponding to no privacy guarantee? In L58, typically one chooses $\delta$ to be much less than $|D|^{-1}$, but the authors state $\delta=o(|D|)$. Incomplete list of small errors: L70 add -> adds L72 gradient -> gradients L89 vector the -> vector of the Alg 2 line 2, missing space L150 extra parens ((x_1, y_1)

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The proposed method is implicitly stated to only apply to image datasets, but this limitation is not openly addressed by the authors. Similar work used for benchmarking is not limited in the same way. The paper focuses on the utility of private models, but does not mention that private training can have negative effects other than reducing utility (other than in a brief Appendix B without references to prior work). For instance, it is well-known that DP-SGD usually exacerbates the biases in models making them more unfair [A][B][C]. Does data augmentation affect fairness? What are the biases introduced from the synthetic data? Robustness is another topic aligned with Trustworthy ML, and researchers often introduce augmented data during training to improve it. It is possible that the proposed method has positive effects on robustness, but the authors have not addressed this direction despite the significant literature on the intersection of privacy and robustness. [A] Bagdasaryan et al. 'Differential privacy has disparate impact on model accuracy' NeurIPS 2019 [B] Xu et al. 'Removing disparate impact on model accuracy in differentially private stochastic gradient descent' ACM SIGKDD 2021 [C] Esipova et al. 'Disparate impact in differential privacy from gradient misalignment' ICLR 2023 ---- Summary of Discussion: I have read all reviews and rebuttals for this submission. The authors responded to all points of my review. My rating was maintained at 5 as some of my original points remain as concerns, including the originality, and quality (lack of details of reproduction).

Reviewer GjeG6/10 · confidence 4/52023-07-06

Summary

The paper studies mixup data augmentation for differentially private (DP) machine learning. The traditional mixup data augmentation requires multiple samples. Therefore, applying them in DP model training is not straightforward. The paper first shows that a naive way of implementing it with micro-batches of size 2 gives poor results. To improve the performance, the paper proposes two new approaches. The first approach mixes up a sample and its own augmented version. The second approach mixes up a sample with generated samples from diffusion models. Experiments show that the proposed approaches improve over self-augmentation.

Strengths

* The paper is well-written. * Mixup data augmentation is very useful in the non-DP world. However, incorporating that into DP training is non-trivial. The paper is a valuable step towards bringing the benefit of mixup data augmentation to DP training.

Weaknesses

* I have some doubts about whether the experimental comparison is fair. * More ablation studies are needed. See "questions" for the details.

Questions

My major concern is the experimental comparison between self-aug and the proposed mixup variants in Tables 2 and 3. * In the experiments of Table 2, did we keep the number of augmented samples the same? To make it more accurate, let's consider Mixup-SelfAug and use the notations in Section 3.2. For a fair comparison, we need to make sure that the "k" utilized in self-aug equals "k+k'" used in Mixup-SelfAug. (The motivation is to keep the computation cost roughly the same.) * Similarly, for Table 3, did we keep the number of augmented samples the same? Other questions: * It is better to show how k, k', and k'' influence the performance of Mixup-SelfAug and Mixup-Diffusion. * Is there any benefit of combining Mixup-SelfAug and Mixup-Diffusion together? I will update the score according to the answers to the above questions. Other minor issues: * Figure 2: I do not fully understand the figure. It says "The histogram shows average norm of gradients.", but I do not get which values are averaged over. Is the process (1) for each sample, computing the average norms of gradients across all iterations, and (2) plotting the histogram of the above values of all samples? Similarly, how are the lines computed? * Line 138: "however" -> ". However" * Line 150: "((x_1,y_1)" -> "(x_1, y_1)" * Line 249-254: The information in this paragraph is also covered in the method section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Limitations

The paper did not discuss the limitation or the potential negative societal impact.

Reviewer ayCk4/10 · confidence 4/52023-07-06

Summary

This paper proposes a data augmentation technique for differentially private deep learning using mixup regularization. Mixup is a popular augmentation technique which involves taking linear combinations of training samples to create new samples. However, such a technique cannot be directly applied to differentially private training, as the sensitivity of the private algorithm increases rendering it to be un-useful. Towards this end the authors propose two methods for private mixup augmentation. First, is mixup with self augmentation which involves mixing samples obtained from augmentation of a single sample, and secondly, mixing training samples with samples generated from diffusion models. The paper provides empirical results to support their claim.

Strengths

Mixup is an interesting technique, and exploring it for differentially private training is important. The technique is simple to implement in practice with easy to prove theoretical guarantees. Method shows improved empirical performance on the presented datasets.

Weaknesses

The proposed method isnt too different from the method proposed by De et al which is cited in the paper. Its a minor augmentation to the augmentation they proposed. Pre-trained models are trained on large datasets but fine-tuning is performed only on toy datasets. Needs more empirical evidence on practical datasets (for example, oxfordpets, flowers, and so on)

Questions

For diffusion model based experiments the pre-trained model is trained on 2B images which is likely to contain the entire fine-tuning sets. Does this seem like a reliable public pre-training to choose? Can you increase the number of samples to mixup from 2? How does the performance change by increasing the number of samples in the convex combination? What should be the nature of the distribution of samples that are used during mixup? Can out of distribution samples be mixed to obtain meaningful results?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

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

2 fair

Contribution

2 fair

Limitations

The method is only restricted to self augmentations or augmentations with publicly available data. In private training having different gradients aligned when training with DP-SGD is important to reduce the effect of the gaussian noise added to ensure DP. Will it be possible to construct a mixup involving multiple samples (rather than self-mixup or diffusion model based) so that the gradients during training are more aligned, which thereby increase the signal-to-noise ratio?

Reviewer sSKD2023-08-14

Response to rebuttals

I have read all the reviews and rebuttals. 6. I was looking at the WRN-40-4 rather than WRN-16-4 in [10], so this is clear now. I was unclear about what "Baseline" referred to in Table 1, this could be made more explicit in the work. However, the SelfAug's reported performance of 79.5% [10] is significantly different from your number, 78.7% in Table 2, given that Mixup-SelfAug achieves 79.8%. At this level, the approaches are within one standard deviation on performance. 10. Does Fig 2 not show that the average norm is reduced? If the distribution is more concentrated around zero, wouldn't the average norm be reduced?

Authorsrebuttal2023-08-17

Thank you for your insights and feedback. 1. We will clarify the meaning of “Baseline” in Table1 of the revised paper. 2. To ensure a fair comparison, we employ Sander et al.’s (ICML 2023, https://openreview.net/pdf?id=DIkGgI9baJ) reproduction of [10] for both SelfAug and Mixup-SelfAug. 3. Fig 2. does indeed show that the average norm is reduced – there's a notable reduction from approximately 0.009 to 0.0004 for CIFAR-10 and 0.001 to 0.0003 for Caltech256.

Authorsrebuttal2023-08-20

For completeness, we train WRN-16-4 on CIFAR10 using Self-Aug and our proposed Mixup-SelfAug using [10]’s official JAX code (https://github.com/deepmind/jax_privacy). As in [10], we repeat each experiment 5 times and report median test accuracy in the table below. The results are consistent with those presented in our paper – Mixup-SelfAug outperforms Self-Aug for different privacy budgets. | Method | $\varepsilon = 1$ | $\varepsilon=8$ | |--------------|------------------|------------------| | Self-Aug | 56.28%(±0.30%) | 79.42%(±0.14%) | | Mixup-SelfAug| 57.05%(±0.35%) | 80.02%(±0.18%) |

Reviewer GjeG2023-08-16

The rebuttal fully addressed my concerns. Therefore, I increase the score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC