acceptodds
Under review as a conference paper at ICLR 2027

Scalable Differentially Private Compression via Diffusion and Stochastic Codes

Abstract

The ever-increasing collection of personal data has created mounting pressure to develop technologies that protect sensitive aspects of individual identity. Differential privacy (DP) provides a principled framework with strong formal guarantees and has already achieved practical success. However, releasing high-dimensional data, such as images, has remained elusive: releasing uncompressed privatized data requires significant storage. At the same time, no effective data compression scheme exists that can compress high-resolution data with privacy guarantees. We address this challenge with DP-DiPP, a compression pipeline that combines stochastic codes with diffusion models. DP-DiPP is highly flexible: the practitioner has direct control over the compression rate-privacy-utility tradeoff. As the theoretical backbone, we extend the Poisson private representation (PPR) of (Liu et al., 2024) to encode the outputs of privacy mechanisms. We then combine it with DiffC, a diffusion-based lossy data compression method, to obtain a differentially private image compressor. Our experiments on privatized image classification on CIFAR-10 demonstrate that DP-DiPP significantly outperforms the baseline, achieving a 10-30 times better compression while retaining comparable privacy guarantees and utility.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion model, differential privacy, image compression, relative entropy coding, channel simulation, DiffC

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_2gCSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the joint problem of image privacy protection and compression under the Local Differential Privacy (LDP) framework. DP-DiPP integrates two components: (i) an approximate steplimited, candidatebounded Poisson Private Representation (PPR) method, which is claimed to provide a 2αε-LDP guarantee for encoding ε-LDP mechanisms; (ii) a modified diffusionbased compressor (DiffC), in which Laplace channels with momentmatching techniques replace conventional Gaussian reversestep targets to achieve provable pure LDP guarantees. Experiments are conducted on the CIFAR10 dataset with ResNet56 and the 512×512 Shenzhen chest X-ray dataset with DenseNet121. The baselines for comparison include the PNGcompressed Laplacemechanism, LDPSlicing, and Gaussian DiffC. Substantial bitrate improvements are reported across plotted privacyutility operating points. Variance analyses over three experimental runs are provided in Appendix H. The paper proves a general result for steplimited PPR (Theorem 3.1), with further implementation details given in Appendices B through J.

Critical strengths
  1. This paper addresses the challenge that the privacy guarantees of a mechanism cannot automatically extend to its stochastic encoding. The authors propose a step‑limited PPR approach, replace Gaussian diffusion targets with Laplace distributions, and develop a diffusion‑based compression method with formal LDP guarantees. Theorem 3.1 together with the privacy analyses in Sections 3‑4 and Appendices C‑D provide theoretical underpinnings for this design, making the contributions of this work go well beyond empirical improvements over existing compression methods.
  2. Experiments on CIFAR‑10 and high‑resolution chest X‑ray images demonstrate a substantial reduction in bit rate compared with the “privatize‑then‑compress” baseline and LDP‑Slicing. The paper additionally includes implementation details, experimental setups, and variability results across three runs in the appendices (Sections 5.1‑5.4; Appendices G‑J). These results suggest that DP‑DiPP is a promising approach for privacy‑preserving image compression.
Critical weaknesses
  1. Theorem 3.1 requires n≥2^{l_α(x)/β}, where the sampling budget depends on the KL‑divergence of each individual input. However, Appendix J.2 employs fixed per‑step KL budgets estimated from a calibration dataset. It remains unclear whether these fixed budgets satisfy the per‑image approximation guarantees, especially for images with large KL‑divergence values. Furthermore, there appears to be a gap in the proof given in Appendix C: Equation (17) upper‑bounds E[logKn], whereas Equation (18) applies Markov’s inequality to the untruncated index K. The authors should clarify this step and elaborate on the relationship between the theoretical approximation guarantees and the practically implemented encoder.
  2. Although the paper highlights fixed‑budget encoding and GPU parallelization, it does not report practical encoding/decoding wall‑clock time for high‑resolution images, peak GPU memory consumption, or the number of required PPR proposals. Appendix B discusses the efficiency of the original DiffC implementation but does not directly evaluate the computational cost of DP‑DiPP.
Questions for authors

Q1. Theorem 3.1 provides a total‑variation‑distance upper bound for step‑limited PPR, where the required sampling budget depends on the KL divergence between the target and proposal distributions. However, Appendix J.2 uses fixed per‑step KL budgets estimated from calibration data. Does this approximation upper bound still hold for images whose actual KL divergence exceeds the estimated average value?

Q2. This paper emphasizes the scalability of DP‑DiPP achieved via step‑limited PPR and GPU parallelization, yet it provides limited results regarding computational efficiency. Could the authors report the encoding‑decoding wall‑clock time and peak GPU memory usage, especially for high‑resolution chest X‑ray images, and compare them against vanilla DiffC and other baselines?

6 comments
AuthorsOpenReview ↗

Responses (part 1)

Thank you for your insightful comments, and for taking the time to review the appendix.

There appears to be a gap in the proof in Appendix C: Equation (17) upper‑bounds E[logKn] ...

The apparent gap in the proof of the approximation bound arose from a typographical error, which we have now corrected. With this correction, Equation (17) establishes a valid upper bound on E[logK]\mathbb{E}[log K]. The rest of the proof follows as before.

Theorem3.1 provides a TV distance bound […] where the sampling budget depends on the KL-divergence of each individual input. AppendixJ.2 uses fixed per-step KL […]. Does this bound still hold?

We agree that the fixed KL budgets estimated from calibration data do not necessarily satisfy the sufficient condition for the per-image TV approximation guarantee in Theorem 3.1. When KL is hard-coded, the upper bound on the TV approximation error grows linearly with the actual KL divergence for a fixed sampling budget nn, before being capped at 1. Thus, images with larger KL divergence may have weaker approximation guarantees. However, as brought to our attention by Reviewer 1, the bound already becomes vacuous even in the non-fixed KL budget scenario at the practical sampling budget used in our experiments (n≤216n\leq 2^{16} per chunk), which we now state explicitly in the paper (L361–372).

Nevertheless, as our additional experiments demonstrate, the practical approximation error has only a limited effect on reconstruction quality in the evaluated settings, including images whose KL divergence exceeds the calibrated budget. Importantly, the vacuity of the TV bound does not invalidate the privacy guarantee of step-limited PPR.

The authors should […] elaborate on the relationship between the theoretical approximation guarantees and the practically implemented encoder.

To empirically investigate the effect of using step-limited PPR compared to exact sampling, in Appendix K, we replace every PPR-coded sample with an exact sample of the same reverse-step target on 50 random X-rays, keeping the initial noise, time-steps, and denoiser fixed. Before denoising, PSNR differs by at most 0.036 dB and SSIM by at most 0.0023. After denoising, exact sampling is at most 0.62 dB higher in PSNR, and SSIM differs significantly only at the two largest budgets (ϵ\epsilon= 46.9 and 82, by 0.017 and 0.016, respectively) in favor of exact sampling. The gap does not grow with the number of encoding steps. This is expected from the algorithm's structure: at every step, the target q(xt−1∣xt,x0)q(x_{t-1} | x_t, x_0) is computed from the current state and the original image, so a deviation introduced at one step can be corrected at the next.

We further investigate the effect of hard-coding the KL budget in Appendix L.

We start by comparing each X-ray’s own KL with the hard-coded budget at every timestep transition, and we find that the budget exceeds the image’s total KL for all 662 images at every transition except the last two, where 7.3% and 45.8% of the images exceed it. Since only the two largest privacy budgets use these transitions, we test the worst case, ϵ\epsilon = 82: the 50 images with the most chunks exceeding the budget are re-encoded with their own KL at every step. On these images, the rate increases by 0.0040 ±\pm 0.0007 bpp (0.0636 →\rightarrow 0.0676 bpp). For the denoised images, PSNR increases by 0.16 ±\pm 0.15 dB, and SSIM does not change significantly; for the noisy images, PSNR does not change, and SSIM increases by 0.001. We note that DP-DiPP hard-codes the budget because encoding with the image’s own KL would make the code depend on x0x_0 beyond the certified ϵ\epsilon.

References

[1] Vonderfecht, J.I., Liu, F., ”Lossy compression with pretrained diffusion models,” in ICLR 2025.

AuthorsOpenReview ↗

Responses (part 2)

This paper emphasizes the scalability of DP‑DiPP […] yet it provides limited results regarding computational efficiency. Could the authors report the encoding‑decoding wall‑clock time and peak GPU memory usage […] and compare them against vanilla DiffC and other baselines?

Our claim of scalability mainly refers to the ability to scale the joint compression and privatization of high-dimensional data, which, to our knowledge, has not previously been made operational.

In Appendix O, we report the wall-clock time and GPU utilization of our method and all the baselines for both datasets. For the high-resolution CXR dataset, across privacy levels, DP-DiPP takes 1.1-6.1 s to encode (time increases with ϵ\epsilon) and 7.0-11.1 s to decode and denoise (time decreases with ϵ\epsilon). In comparison, these values are 0.78-4.6 s and 6.5-11.0 s for the DiffC baseline. For the Laplace baseline, encoding and denoising are done on the encoder side and take 0.029-6.2 s (time decreases with ϵ\epsilon), and decoding takes 0.002 s. For LDP-Slicing, encoding takes 0.047-0.07 s, and decoding takes 0.004-0.006 s. For the same dataset, the peak GPU memory usage for DP-DiPP is 6.56 GBs, the same as the DiffC baseline, while for the Laplace baseline (due to the utilization of the same ADM diffusion model for denoising) the value is 6.55 GBs, and for the LDP-Slicing baseline this value is 0.74 GBs.

Thank you again for your time and constructive feedback, and would be grateful if you could kindly reconsider the evaluation in light of the clarified results and strengthened empirical evidence.

Thank you for the detailed response. Some of my concerns have been addressed. However, the theoretical analysis in the appendix has only been supplemented with additional clarifications regarding its limitations, without substantively resolving the underlying issues. Therefore, I will maintain my original score.

Reviewer_9VeYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DP-DiPP, a method that jointly compresses and locally privatizes images by combining diffusion-based lossy compression (DiffC) with stochastic codes. Because standard privatize-then-compress pipelines waste bits encoding high-entropy privacy noise, the authors instead compress the reverse denoising steps of a pretrained diffusion model using a fixed-runtime, step-limited variant of the Poisson Private Representation (PPR). To obtain a pure ε\varepsilon-local differential privacy (ε\varepsilon-LDP) guarantee, they replace the Gaussian denoising targets and proposals at each diffusion step with moment-matched Laplace distributions, incurring roughly a 2×2\times bitrate penalty over non-private diffusion compression while maintaining downstream utility. Evaluated on CIFAR-10 and the Shenzhen chest X-ray dataset, DP-DiPP achieves 10–30×10\text{--}30\times and 110–920×110\text{--}920\times lower bitrates, respectively, than privatize-then-compress baselines at matched downstream classification accuracy and per-pixel ε\varepsilon-LDP budgets.

Critical strengths
  1. Meaningful extension of PPR significantly reduces complexity.
  2. Privacy guarantees seem sound and experimental results compare favorably to baselines.
Critical weaknesses
  1. Applying Local-DP at a per-pixel level is technically sound yet completely vacuous at providing real privacy protection to images. This becomes paradoxical at the intersection of compression and privacy: image compression works by encoding correlations across pixels and grouping many pixels, while the pixel-wise LDP claims are only meaningful if we consider each pixel as an indepdently informed secret.
  2. a background introduction of the stochastic compression formulation and prior works, indepdent from its role in privacy, would be appreciated.
Questions for authors

Please motivate pixel LDP on when it provides non-vacuous privacy protection.

1 comment
AuthorsOpenReview ↗

Responeses

Thank you for your review. Please let us know if our answers below address your concerns; we also welcome any further actionable feedback that we could fix that would help you lean more towards the acceptance of our paper.

Applying Local-DP at a per-pixel level is technically sound yet completely vacuous at providing real privacy protection to images…

Thank you for raising this concern. We agree that a per-pixel LDP guarantee alone may not provide strong protection for sensitive information distributed across an entire image. However, we would like to clarify the following points.

First, our privacy framework is not restricted to protecting individual pixels. The per-pixel guarantee also induces a formal image-level guarantee through standard privacy composition. It does not require pixels to be statistically independent and remains valid when the image contains correlations across pixels.

We acknowledge that obtaining better image-level privacy guarantees by exploiting the structure of our mechanism or proposing better composition for high-resolution images is an interesting direction for future research.

Second, step-limited PPR compresses any pure ϵ\epsilon-LDP mechanism under any adjacency and metric. We report per-pixel privacy as it provides a common basis for comparison across different image resolutions, analogous to the use of bits per pixel (bpp) in image compression, and with the existing pixel-level LDP methods (Cao et al., 2026). It allows us to characterize privacy under the same single-pixel adjacency notion for both low- and high-resolution images, while the corresponding full-image guarantee can be obtained through composition.

Finally, our choice of DP is motivated by its rigorous and quantifiable privacy guarantees. Although stronger or more application-specific notions of image privacy may be desirable, DP provides a well-established framework for analyzing privacy and its tradeoff with utility and communication cost. Our primary contribution is to improve the compression–privacy–utility tradeoff within this established framework, rather than to claim that pixel-level DP alone fully captures all aspects of semantic or image-level privacy.

Thank you again for highlighting this important distinction. We clarified the relationship between pixel-level and image-level privacy, together with the limitations of the composition bound, in the revised manuscript (L531-535).

a background introduction of the stochastic compression formulation and prior works, indepdent from its role in privacy, would be appreciated.

We appreciate this concern, and we expanded Section 2 to introduce stochastic compression (L101-116). However, please note that the brevity was not without reason: due to the complex nature of our contribution, the background section was already 3 pages long.

References

[1] Cao, Y., Li, C., He, W., ”LDP-Slicing: Local differential privacy for images via randomized bit-plane slicing,” in CVPR 2026.

Reviewer_6FkZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper investigates joint lossy compression and ldp for images by combining diffusion compression with a step-limited variant of the Poisson private representation (PPR) stochastic code. To get pure epsilon-LDP at each reverse step (in pixel space in their experiments), the authors replace Gaussian denoising transitions with moment-matched Laplace distributions. The question the paper asks is intersting, but a practical threat model, a huge gap between the theoretical bound and practice, missing runtime numbers, and several baseline/figure issues need to be addressed. I will consider increasing my score if at least some of these are provided

Strengths and Weaknesses

  • Generally the paper is well-written; the motivation in Sec. 2 on why privatize-then-compress wastes bits on noise is clear, and the Taylor expansion in App. F explaining the 2x bitrate factor from Laplace moment-matching is quite nice.

  • L283-L294 (Theorem 3.1) and appendix b and c -> The TV approximation bound n >= 2^(l_alpha(x) / beta) is vacuous in the regime where the method actually runs. In App. B you mention transitions are split into chunks of ~16 bits of KL. With alpha = 2, even a loose TV bound of beta = 0.1 requires n >= 2^196 candidates per chunk, whereas realistic GPU budgets (e.g., n = 2^16) give beta > 1. What is the actual budget n used in your experiments, and what happens when a test image's KL exceeds the hard-coded calibration average (App. J.2)?

  • L313-L323 & L480: How meaningful is a per-pixel LDP threat model when per-pixel epsilon goes up to 82 on 512x512 X-rays and 256–1000 on CIFAR-10 (and scales by HxW for the whole image)? You mention in L480 that pixel-level guarantees do not prevent global content recognition, yet there are zero visual examples of the images in the paper. Showing visulizations of the reconstructions (and PSNR/SSIM) is specially important on Shenzhen CXR, where you use an ImageNet-trained ADM generator with a fixed class label (L915) to denoise medical X-rays.

  • Copying 1-channel greyscale X-rays to C = 3 channels before privatization inflates your per-pixel epsilon by 3x and artificially triples the bitrate of the PNG-compressed Laplace baseline compared to privatizing the native 1-channel image. Also, L401 says the Laplace baseline is denoised with the diffusion model before PNG compression, while L423 says it is just clamped to [0, 255] and PNG-compressed—which one was actually run?

  • The title and contributions emphasize a scalable, fixed-runtime parallelizable PPR algorithm with a custom CUDA kernel, but there is no wall-clock runtime or latency reported anywhere, if not provided authors at least nee to change the title.

  • Arguing that DP-DiPP has the same empirical privacy as the 4x-tighter Laplace baseline because an obvious empirical attack only attacks the decoded output directly contradicts Sec. 2.3 (Eq. 2), where the whole point of using PPR is that the attacker observes the compressed bitstream enc(X, Z) directly.

  • Clarification of Fig. 1, 2, and 4 is necessary, caption notes that Gaussian DiffC does not have an LDP certificate (epsilon = inf), so why is it plotted as a curve along the finite epsilon-LDP certificate x-axis by giving it the same x-coordinates as DP-DiPP?

  • In Fig. 1 and 4, the middle/bottom-right x-axes cut off at 300 even though the largest CIFAR-10 point (epsilon = 256 with 2 * alpha = 4) is at 1024, leaving the curves hanging off the border. Am I missing something here?

Minor

Minor: L018: the the and L097: a system stem L360-L367: n is used for both the PPR candidate budget (L360) and the number of diffusion steps (L367). L647 & L650: in L647 > k should be > n, and L650 writes K_n^alpha instead of the exact index K when applying Theorem 4.3 of liu ? L872 vs L874: L874 drops the sqrt(2) factor that appears in L872 and Eq. 15—which one was used in the code?

Critical strengths

Please see the Summary

Critical weaknesses

Please see the Summary

Questions for authors

Please see the Summary

3 comments
AuthorsOpenReview ↗

Responses (part 1)

Thank you for your meticulous and insightful comments, especially for taking the time to review the appendix. We address your concerns below. Please let us know if you have any further questions or concerns. Moreover, we have substantially revised our paper, benefiting from the allowed additional page. All the changes are highlighted in blue.

The TV approximation bound [...] is vacuous… What is the actual budget n used in your experiments

Following [1], each reverse step is split into chunks with a 14-bit KL budget and 2 bits of headroom, so n≤216n \leq 2^{16} per chunk. We agree that for this n, the bound in (11) becomes vacuous. We now state this explicitly in the paper (L361-372). However, the ℓα(X)/β\ell_\alpha(X)/\beta scaling of the sample complexity in our bound on the TV error of the sampler is probably the best one can do, please refer to the achievability and converse results of sample complexity in [2, Theorems 5 and 6]. Intuitively, this is because a guarantee on the TV distance is quite strong, so it is somewhat unsurprising that the bound quickly becomes loose.

Nevertheless, as our experiments show, this does not affect the practical usefulness of our scheme, the privacy guarantee, or the rate of step-limited PPR. In addition to the aggregate results, we now include visual examples in Figure 3 in the main text and in Appendix N.

In particular, in Appendix K, we show that step-limited PPR does not significantly affect image quality. On 50 random X-rays, we replace every PPR-coded sample with an exact sample of the same reverse-step target, keeping the initial noise, time-steps, and denoiser fixed. Before denoising, PSNR differs by at most 0.036 dB and SSIM by at most 0.0023. After denoising, exact sampling is at most 0.62 dB higher in PSNR, and SSIM differs significantly only at the two largest budgets (ϵ\epsilon= 46.9 and 82, by 0.017 and 0.016, respectively) in favor of exact sampling. The gap does not grow with the number of encoding steps. This is expected from the algorithm's structure: at every step, the target q(xt−1∣xt,x0)q(x_{t-1} | x_t, x_0) is computed from the current state and the original image, so a deviation introduced at one step is corrected at the next.

What happens when a test image's KL exceeds the hard-coded calibration average

Even in this case, the image quality does not degrade significantly.

Concretely, in Appendix L we investigate the effect of hard-coding the KL budget. Comparing each X-ray’s own KL with the hard-coded budget at every timestep transition, we find that the budget exceeds the image’s total KL for all 662 images at every transition except the last two, where 7.3% and 45.8% of the images exceed it. Since only the two largest privacy budgets use these transitions, we test the worst case, ϵ\epsilon = 82: the 50 images with the most chunks exceeding the budget are re-encoded with their own KL at every step. On these images, the rate increases by 0.0040 ±\pm 0.0007 bpp (0.0636 →\rightarrow 0.0676 bpp). For the denoised images, PSNR increases by 0.16 ±\pm 0.15 dB, and SSIM does not change significantly; for the noisy images, PSNR does not change, and SSIM increases by 0.001. We note that DP-DiPP hard-codes the budget because encoding with the image’s own KL would make the code depend on x0x_0 beyond the certified ϵ\epsilon.

References

[1] Vonderfecht, J.I., Liu, F., ”Lossy compression with pretrained diffusion models,” in ICLR 2025.

[2] Block, A., Polyanskiy, Y., ”The Sample complexity of approximate rejection sampling with applications to smoothed online learning,” in COLT 2023.

AuthorsOpenReview ↗

Responses (part 2)

How meaningful is a per-pixel LDP threat model when per-pixel epsilon goes up to 82…

Here we clarify this point from two angles.

  • First, for a per-pixel epsilon that goes up to 82, we agree that this could look unconventional at first glance. However, we would like to point out that the DP guarantee in high-dimensional images can mean something totally different from what people are used to in the DP literature (which was usually for low-dimensional numerical data). To the best of the authors’ knowledge, it was not until this year, by LDP-Slicing (Cao et al., 2026 CVPR), that pixel-level LDP for images became meaningful. Therefore, we argue that a large epsilon could be acceptable when people consider image-level DP, though this is a very active and recent area, and people are still exploring the meaning and use of DP in high-resolution images. We also have medical image examples in the revised paper, please refer to Figure 3.

  • Second, for the DP guarantee of an image, as also pointed out by the reviewer, one would need to use composition of per-pixel DP to derive a per-image DP guarantee. Therefore, our privacy framework is not restricted to protecting individual pixels, and it does not require pixels to be statistically independent and remains valid when the image contains correlations across pixels. Moreover, we report per-pixel privacy as it provides a common basis for comparison across different image resolutions, analogous to the use of bits per pixel (bpp) in image compression, and with the existing pixel-level LDP methods (Cao et al., 2026 CVPR).

Importantly, the use of DP is motivated by its rigorous and quantifiable privacy guarantees. Although stronger or more application-specific notions of image privacy may be desirable, DP provides a well-established framework for analyzing privacy and its tradeoff with utility and communication cost. We acknowledge that obtaining better image-level privacy guarantees by exploiting the structure of our mechanism or proposing better composition for high-resolution images is an interesting direction for future research.

Moreover, in Appendix N, we provide the PSNR/SSIM values for our method and all other baselines for all the privacy levels. Also, we provide the noisy and reconstructed images for these privacy levels for 3 512 ×\times 512 CXR images.

Copying 1-channel greyscale X-rays to C = 3 channels before privatization

This was an oversight on our part. Thank you for pointing this out; we have now fixed this in the revised manuscript. We provide the new results alongside the previous results in Appendix M.

As you suggested, we changed our Laplace privacy mechanism baseline to add noise calibrated to the privacy level in a single channel. If we denoise before release, since the denoiser expects 3 channels, we first copy the single-channel values to three channels, then denoise (as detailed in revised Section 5.1), and finally average the channels in the denoised image to recover a single-channel image; then we compress the greyscale image as a PNG. For the TB classification experiments, we again copied the single-channel image to three channels to match our classification architecture, and then report the classifier trained with and tested on three-channel images. We provide the new results alongside the previous results in Appendix M, and we have updated our claims and figures in the main paper accordingly.

For the previous Laplace baseline, we reported the compression rate and utility for the denoised Laplace baseline. Thank you for bringing L419 to our attention; we made the necessary corrections in Section 5.1 of the revised manuscript.

no wall-clock runtime or latency reported anywhere

Our claim of scalability mainly refers to the ability to scale the joint compression and privatization of high-dimensional data, which, to our knowledge, has not previously been made operational.

In Appendix O, we report the wall-clock time and GPU utilization of our method and all the baselines for both datasets. Across privacy levels, DP-DiPP takes 1.1-6.1 s to encode (time increases with ϵ\epsilon) and 7.0-11.1 s to decode and denoise (time decreases with ϵ\epsilon). In comparison, these values are 0.78-4.6 s and 6.5-11.0 s for the DiffC baseline. For the Laplace baseline, encoding and denoising are done on the encoder side and take 0.029-6.2 s (time decreases with ϵ\epsilon), and decoding takes 0.002 s. For LDP-Slicing, encoding takes 0.047-0.07 s, and decoding takes 0.004-0.006 s.

AuthorsOpenReview ↗

Responses (part 3)

Arguing that DP-DiPP has the same empirical privacy as the 4x-tighter Laplace baseline …

Thanks for pointing out that our empirical attack does not capture the worst case. We generally agree: the strongest adversary would attack the compressed code directly. This is exactly the case our PPR analysis covers. Its privacy guarantee holds for any adversary observing the code, so the formal guarantee does not rely on the empirical attack being optimal.

The empirical attack in the paper is meant as a good-faith effort to complement the bound by showing privacy in practice, not to replace it. To our knowledge, attacks that target the outputs of stochastic compression of privacy mechanisms have not yet been studied, and it is not clear to us how they could be implemented. We see designing them as a promising direction for future work. We made this distinction explicit in L520-523.

Clarification of Fig. 1, 2, and 4 is necessary

We changed the figure captions to reflect that the DiffC baseline doesn’t provide a formal privacy guarantee, but is included to show that using PPR instead of PFR and a Laplace channel instead of a Gaussian channel for denoising doesn’t degrade downstream performance when the timesteps are kept fixed. We previously cropped the image to focus on the region of interest; we now adjusted it to include all points.

Typos

We corrected all the typos. In the code, we use the 2\sqrt{2} version.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their constructive reviews. Below, we summarize the main revisions we made in response. All changes are highlighted in blue in the revised PDF and are referenced by section in the individual responses.

  1. Stronger Laplace baseline on X-rays (Sec. 5.1; App. M). Noise is now added to the single grey channel instead of three copied channels, and the rate is that of the denoised 1-channel PNG.

  2. Approximation bound vs. practice (L361–372; App. C, K, L). We state that the TV bound is vacuous at our budget (n=216n = 2^{16}) and fixed a typo in its proof (Eq. 17). We compare step-limited PPR with exact sampling, and the hard-coded KL budget with each image's own KL.

  3. Visual examples and image quality (Fig. 3; App. N). Releases of all methods at every privacy level, with PSNR/SSIM over all 662 X-rays.

  4. Runtime and memory (App. O). Per 512×512 X-ray, the sender side (encoding, plus denoising for Laplace) takes 1.1–6.1 s for DP-DiPP, 0.8–4.6 s for DiffC, 0.3–6.2 s for Laplace and 0.05–0.07 s for LDP-Slicing; the receiver side (decoding, plus denoising for DP-DiPP and DiffC) takes 7.0–11.1 s, 6.5–11.0 s, 0.002 s and 0.005 s.

  5. Presentation (Sec. 2.2). New background on stochastic compression.

Overall, the conclusions are unchanged. We address the remaining points in the individual responses.

We warmly welcome any further questions, comments, or discussion from the reviewers and the Area Chair. If the reviewers find our responses and revisions satisfactory, we would be very grateful if they could kindly reconsider their assessments and ratings in light of the clarifications, corrections, and new results.

AuthorsOpenReview ↗

Follow-up on Author Responses and Revisions

Dear Reviewers and Area Chairs,

As the discussion period will approach its end in about one day, we would like to kindly follow up on our responses.

We have provided detailed responses to the reviewers' concerns and substantially revised the manuscript based on your constructive feedback, with all changes highlighted in blue. Our main additions include:

  1. Stronger baselines: Updated the Laplace baseline on chest X-rays by privatizing the original single-channel images and reporting the bitrate of the denoised grayscale PNG (Sec. 5.1; App. M).

  2. Theoretical clarifications and analyses: Corrected the proof of Theorem 3.1 and clarified about the TV bound. On 50 X-rays, exact sampling changes pre-denoising PSNR by at most 0.036 dB. Using image-specific instead of fixed KL budgets increases bitrate by only 0.0040 ± 0.0007 bpp, with a denoised PSNR improvement of 0.16 ± 0.15 dB (Apps. K–L).

  3. Visualizations and image quality: Added visual comparisons across all privacy levels (Fig. 3), together with PSNR/SSIM evaluations over all 662 Shenzhen X-rays (App. N).

  4. Runtime and memory: Added comprehensive computational measurements (App. O). For 512×512 X-rays, DP-DiPP requires 1.1–6.1 s for encoding and 7.0–11.1 s for decoding and denoising, with 6.56 GB peak GPU memory, comparable to DiffC.

We particularly thank Reviewer 2gCS for the constructive discussion and increased rating from 2 to 3. We greatly appreciate this encouragement and would welcome further feedback from the other reviewers and Area Chairs.

We sincerely hope these revisions adequately address all the concerns. We would be grateful if other reviewers could kindly reconsider your assessments in light of the clarifications, corrections, and additional experimental evidence.

Thank you again for your time, thoughtful feedback, and consideration!

The Authors

AuthorsOpenReview ↗

One More Gental Follow-up

Dear Reviewers and Area Chairs,

As the discussion period approaches its end (in 1 hours), we would like to kindly follow up on our responses.

We have provided detailed responses to all reviewers' concerns and substantially revised the manuscript based on your constructive feedback, with the revisions highlighted in blue.

We sincerely hope these revisions have addressed your concerns. We would be grateful if you could kindly reconsider your assessments based on our strengthened results, additional experiments, and substantially revised presentation.

Thank you again for your time, thoughtful feedback, and consideration!

The Authors

AuthorsOpenReview ↗

Summary of Contributions and Rebuttal Revisions for the Area Chair

Dear Area Chair,

As the discussion period concludes, we would like to briefly summarize our main contributions and the additional evidence provided during the rebuttal to facilitate your assessment.

1. Main Contributions and Novelty

Our central contribution is to enable practical joint compression and privatization of high-resolution images with formal differential privacy guarantees, addressing the fundamental inefficiency of conventional privatize-then-compress approaches.

  • Novel stochastic coding algorithm and theoretical guarantees: We develop step-limited Poisson Private Representation (PPR), a fixed-budget, parallelizable stochastic code that retains the original PPR's 2αϵ2\alpha\epsilon-LDP certificate and near-optimal expected codelength bound while providing a theoretical characterization of its sampling approximation error (Theorem 3.1). This contribution is general and independent of diffusion models or image compression.
  • Differentially private diffusion-based compression: We identify each reverse diffusion step as a privacy mechanism and introduce moment-matched Laplace transitions in place of Gaussian transitions, enabling pure LDP guarantees while retaining useful denoising performance.
  • Scalable joint privacy–rate–utility framework: By integrating these two contributions, we develop DP-DiPP, which avoids spending bits on encoding privacy noise. Our experiments demonstrate 9–830× and 48–290× lower bitrates than privatize-then-compress baselines on CIFAR-10 and high-resolution chest X-rays, respectively, at matched pixel-level privacy certificates, with comparable downstream utility across the evaluated operating points.

2. Main Additions and Clarifications During Rebuttal

We have substantially revised the manuscript, with changes highlighted in blue, and conducted extensive additional experiments.

  • Theoretical clarifications: Corrected the proof of Theorem 3.1 and clarified that the TV approximation bound is vacuous at the practical sampling budget, but importantly, this does not invalidate the formal LDP guarantee or expected codelength bound.
  • Approximation quality and fixed KL budgets: Compared step-limited PPR with exact sampling on 50 chest X-rays, showing a pre-denoising PSNR difference of at most 0.036 dB. Using image-specific rather than fixed KL budgets increases bitrate by only 0.0040 bpp, with a modest PSNR improvement of 0.16 dB (Appendices K–L).
  • Stronger baselines and image quality: Corrected the Laplace baseline to privatize and compress native single-channel grayscale X-rays. Added visual comparisons across all privacy levels and comprehensive PSNR/SSIM evaluations over all 662 Shenzhen X-rays. Our main empirical conclusions remain unchanged.
  • Computational efficiency: Added comprehensive runtime and memory measurements. For 512×512 chest X-rays, DP-DiPP requires 1.1–6.1 seconds for encoding and 7.0–11.1 seconds for decoding and denoising, with 6.56 GB peak GPU memory, comparable to non-private DiffC.
  • Privacy interpretation: Clarified the relationship between pixel-level and image-level privacy, the limitations of composition, and the distinction between formal bitstream-level privacy guarantees and empirical attacks on reconstructed images.

We particularly appreciate Reviewer 2gCS's acknowledgment that our theoretical corrections and additional experiments have addressed most of their concerns, that the approximation-bound limitation does not invalidate our formal privacy guarantee, and their increased rating from 2 to 3.

We respectfully emphasize that our paper makes two distinct yet complementary contributions: a general stochastic coding algorithm with rigorous privacy guarantees, and a practical framework that makes joint differentially private compression feasible for high-resolution images. The substantial bitrate reductions, strengthened by the additional experiments, demonstrate the practical significance of this framework.

We sincerely thank the Area Chair and all reviewers for their time, constructive feedback, and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.