acceptodds
Under review as a conference paper at ICLR 2027

Privacy Accounting for Empirical Denoisers

Abstract

Diffusion models have emerged as the foundation of modern generative systems, yet their high memorization capacity raises privacy concerns. While differentially private (DP) training provides formal guarantees, it comes at a cost in sample quality. In this work, we study a complementary question: how much privacy loss is incurred during the sampling process itself? Directly analyzing modern neural denoisers is challenging because their sensitivity to a single training example is generally intractable. To overcome this challenge, we construct and analyze a clipped privacy-accountable empirical denoiser, a training-free estimator of the Bayes-optimal denoiser, whose per-step sensitivity can be computed explicitly, allowing each denoising step to be a Gaussian mechanism. This perspective enables privacy accounting under Gaussian DP when each training example contributes to each denoising step, and under -DP trade-off functions when each generated sample uses a single randomly subsampled block of the training data. We report every guarantee as a certified upper bound from a privacy loss distribution accountant, and use a cumulant-based Edgeworth approximation as a fast diagnostic, characterizing where it is reliable and where it becomes anti-conservative. We also analyze the privacy loss along the denoising trajectory and show that it grows steeply near the critical window in which semantic content is formed. This observation motivates hybrid samplers that use the private empirical denoiser through the formative steps where it remains a faithful proxy for the neural denoiser, then switch to a public neural denoiser for the remaining refinement. Our results on CIFAR-10 and CelebA-64 show that the privacy of diffusion models depends not only on the training procedure, but also on sampling-time design choices such as subsampling, stopping time, and hybrid denoising.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
differential privacy, diffusion-based sampling, privacy accounting, gaussian differential privacy, empirical denoiser

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KTtXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes a training-free private generation mechanism via diffusion. It relies on the standard diffusion SDE involving a private empirical denoiser instead of the standard (privately or non-privately trained) empirical denoisers. In this setting, privacy arises from the Gaussian mechanism naturally induced from the diffusion term in the SDE. To scale this method without catastrophic privacy loss accumulation, the paper introduces two tricks: a diffusion model pretrained on a public dataset used instead of the private denoiser for lower noise levels, and a batch subsampling per generated sample to compute the empirical denoiser. The authors then compute the privacy guarantees of such procedure and test it w.r.t. standard DP diffusion models.

Critical strengths
  1. The paper provides an in-depth study of the privacy incurred by the stochastic sampling procedure of diffusion models. The general approach is original and the strong privacy guarantees yield competitive generation results w.r.t. the state of the art. While the impact on future work on DP diffusion is hard to assess, it may inspire other works to improve existing DP-trained diffusion models.
  2. The paper is, for the most part, clear and easy to read. All ideas and hypotheses are well exposed. Up to the extent of my assessment -- I could not review the appendices in-depth, as my technical expertise in differential privacy remains limited --, the theoretical results presented in the paper appear to be sound.
Critical weaknesses

The paper is limited by two important weaknesses.

Motivation and consequences of privatizing the sampling procedure only

  1. The paper provides little and insufficient motivation to privatize the sampling procedure of diffusion only. The statement that "regulators see only generated samples" is incorrect; for example, the EU supervising regulatory body for personal data protection specifically concluded that a model may not be considered anonymous if personal data could be retrieved from model weights (https://www.edpb.europa.eu/news/edpb-opinion-on-ai-models-gdpr-principles-support-responsible-ai_en). Further, even users may have more advanced access to the model, e.g. deployers of a model or even individuals working with open-weight models. The privatization of the sampling procedure may have other motivations, but they are not developed in this paper.
  2. The presented method has a crucial weakness that is discussed and partially alleviated in the paper but that is not recognized as a blocker for scaling and not analyzed thoroughly: the privacy guarantees heavily depend on the number of generated samples. In many application cases of generative models, the number of samples they are tasked to produce is large and may be much larger than the size of their training dataset. The proposed method is therefore impractical.
  3. In particular, the paper should include an analysis of the evolution of the privacy guarantees w.r.t. the number of generated samples.

Experimental relevance

  1. The trick leveraging a denoiser pretrained on public data is clever, but confuses the interpretation of the experimental results. It could be used for all other baselines, which may improve their results. Moreover, the paper should verify the FID of the generation process using this public denoiser only, to ensure that it is not even better than the proposed method itself.
  2. The proposed method will not scale well with the dataset size as it relies on the empirical denoiser. In my opinion, this is not a weakness of the paper since related works also have scaling issues preventing their application at scale. Nonetheless, this should be discussed in the paper along with an analysis of the computational complexity incurred by the empirical denoiser.

[Minor] Clarity

The clarity of some parts of the paper should be improved, should it be accepted after the reviewing process. In particular:

  • more technical background for Gaussian DP should be included in the paper;
  • the purpose of Section 2.3 is unclear;
  • Proposition 4.1 and its surroundings are overly complex to parse and interpret, especially as it uses some notations that have not been previously defined.
Questions for authors

Given the above assessment, I think that this paper presents an interesting method but is limited by its lack of motivation and some of its empirical validation. For these reasons, I recommend a weak rejection. I remain open to discussing with the authors and other reviewers before providing my final recommendation.

In particular, I may change my assessment primarily if:

  1. the above concern about the motivation for a private sampling procedure and the scaling analysis are adequately addressed in the author's response;
  2. the experimental issues are fixed, notably the clarification of the role of the public denoiser.

Post-rebuttal update

I would like to thank the authors for their thorough response, and to apologize for not having had the time to acknowledge it sooner. My main concerns have been adequately addressed. While I'm not quite sure about the impact of the method itself, I believe the developed ideas are worth sharing and could be leveraged in other works. I am therefore happy to now recommend a weak accept.

2 comments

Thank you for recognizing the originality of the sampling-time analysis, and for identifying the two issues that could change your assessment: the motivation and scaling of a private sampler, and the contribution of the public denoiser. We agree that both needed stronger treatment.

Motivation for privatizing sampling

Thank you for the EDPB reference. You are right that "users and regulators see only generated samples" is incorrect. Model weights can reveal personal data, and deployers or open-weight users can access them directly. We removed the statement and rewrote the motivation.

No network is trained on the private data. The private dataset, D\mathcal D, enters only through the clipped empirical denoiser during sampling. The neural denoiser is trained on public data disjoint from D\mathcal D (Algorithm 1), so there are no private weights to leak. A trusted data holder needs access to D\mathcal D to run the hybrid sampler. Deployers and open-weight users receive only the released samples, the public checkpoint, or a model trained on the released samples, and none of these reveal more about D\mathcal D than the (ε,δ)(\varepsilon, \delta) guarantee allows. The public checkpoint must be fixed independently of D\mathcal D.

Why privatize sampling? Private sampling gives a formal guarantee without training any network on the private data. This avoids private training methods and its cost in sample quality, and one public model can serve many private datasets with no training per dataset. Because each step's sensitivity is explicit, the analysis also shows where privacy loss accumulates along the trajectory, which motivates our hybrid sampler. A model trained on the released samples inherits its guarantee by post-processing.

When DP training is better. Each new sample from D\mathcal D adds privacy cost. If an application needs repeated, fresh access to the private data, a DP-trained model may give a better trade-off. Our method is the right fit when the release is bounded and avoiding private training methods altogether is valuable.

Scaling with the number of outputs

We agree this is a genuine limitation. We compute how the certified bound grows with the number of released images RR, with the same PLD accountant and everything else fixed (CIFAR-10, n=50,000n=50,000 B=25B = 25, q=5×10−4q = 5 \times 10^{-4}, δ=10−5\delta = 10^{-5}, add/remove adjacency). Outputs are generated independently, so per-image quality does not depend on RR, and larger releases need only new accountant runs.

55% handoff62% handoff65% handoff
FID (50k, one seed)45.6729.6625.00
εu\varepsilon_u at R=nR = n0.301.696.29
εu\varepsilon_u at R=10nR = 10n1.055.2817.9
εu\varepsilon_u at R=100nR = 100n3.8321.380.6
Largest RR, ε≤1\varepsilon \le 1461k (9.2n9.2n)10.5k (0.21n0.21n)23
Largest RR, ε≤5\varepsilon \le 57.77M (155n155n)453k (9.1n9.1n)23k (0.47n0.47n)
Largest RR, ε≤10\varepsilon \le 1022.6M (452n452n)1.50M (30n30n)159k (3.2n3.2n)

The cost grows sublinearly in RR. In the regime where the hybrid sampler improves FID, a privacy budget of ε≤5\varepsilon \leq 5 allows about 9n9n sampled images at the 62% handoff and about 0.5n0.5n images at 65%. For ε≤10\varepsilon \leq 10, we can generate about 3n3n samples at 65% handoff. A data holder can therefore fix the handoff and the release size before generating to fit within a certain privacy budget.

Unlimited coverage does not mean unlimited utility. Our outputs, like samples from a DP-trained model, are independent draws from one fixed distribution, so downstream utility eventually saturates. On CIFAR-10, Ghalebikesabi et al. (2023) show that a classifier trained from scratch on their synthetic images keeps improving with more samples (their Figure 3). The same classifier pretrained on ImageNet-32 reaches the same accuracy at 50K synthetic images and gains little accuracy from further samples (their Figure 10). Our setting already assumes public data, so downstream models can be pretrained in the same way. Released samples can also be expanded by post-processing at no privacy cost, as Lin et al. (2024) do for Private Evolution.

  • Ghalebikesabi et al. Differentially Private Diffusion Models Generate Useful Synthetic Images. arXiv 2023.
  • Lin et al. Differentially Private Synthetic Data via Foundation Model APIs 1: Images. ICLR 2024.

Sampling with multiple releases

Released samples do not need to be regenerated to serve more users. Reusing those samples costs no privacy, and any model trained only on the release and public data inherits its guarantee. For CIFAR-10, we fine-tuned a class-conditional latent diffusion model (LDM) (Rombach et al., 2022), pretrained on ImageNet 32×32 with 16×16 latents, on the DP released samples from our headline configuration (εu=6.29\varepsilon_u = 6.29), with labels from a public ImageNet classifier applied to the released images. For CelebA-64, we fine-tuned an unconditional LDM pretrained on FFHQ 256x256 on our release (εu=3.89\varepsilon_u = 3.89). No step reads the private data itself.

FID (50k samples, one seed)CIFAR-10CelebA-64
Our release25.0023.00
Pretrained LDM, no fine-tuning25.9943.99
Attention modules 9–16, high noise only23.01 (t≥650t \ge 650)22.52 (t≥184t \ge 184)
Full UNet, high noise only22.64\textbf{22.64} (t≥650t \ge 650)18.90\textbf{18.90} (t≥184t \ge 184)
Attention modules 9–16, more noise levels30.43 (all tt)28.08 (t≥650t \ge 650)

Fine-tuning improves on the utility of the samples themselves. Training only at high noise works best, which suggests the released samples mainly carry coarse structure while the public prior supplies the fine details. This split is a utility choice, since training on the released samples is free at every noise level. Our method can therefore give a model with unlimited output coverage at the release's budget, without DP training.

Despite better baseline FID, the pretrained ImageNet LDM cannot replace the public denoiser (ImageNet EDM) inside our sampler. Our sampler runs the empirical and public denoisers on the same pixel-space trajectory, while the LDM denoises in latent space at every step. It is still useful given the released samples, as post-processing, to improve the utility at no additional privacy cost.

Contribution of the public denoiser

Public-only sampling gives FID 30.70 ±\pm 0.17 (CIFAR-10) and 49.26 ±\pm 0.09 (CelebA-64), against 24.88 ±\pm 0.10 and 22.03 ±\pm 0.10 for our hybrid at the headline setting (three seeds). The private block thus improves FID by 5.82 and 27.23 compared to the public-only baseline. On CIFAR-10, the utility gains appear from the 62% handoff onward, and on CelebA-64 it holds across the whole sweep (see our response to gMB8).

Comparison fairness

Five of the six baselines also use public ImageNet data. DP-Diffusion (Ghalebikesabi et al., 2023) and DP-Promise (Wang et al., 2024) pretrain a diffusion model on ImageNet and fine-tune it on the private data. DP-LDM (Liu et al., 2024) fine-tunes only the attention modules of an ImageNet LDM with DP-SGD. DP-API (Lin et al., 2024) queries an ImageNet diffusion model through its API and, like ours, trains no network on the private data. DP-MEPF (Harder et al., 2023) uses an ImageNet feature extractor. Only DP-DM (Dockhorn et al., 2023) uses no public data. Table 2 uses certified bounds below ε=10\varepsilon = 10 and marks the comparison as contextual.

Computational scaling

For RR outputs, KK private steps and dd-dimensional images, the empirical denoiser costs O(RKnd)O(RKnd) with full batches and O(RKBd)O(RKBd) with output-wise subsampling, on top of public inference. Space complexity is O(Bd)O(Bd) per trajectory, plus O(nd)O(nd) for the dataset. Per-output cost is independent of nn at fixed BB, but grows linearly with nn at fixed q=B/nq = B/n. A DP-SGD model instead pays a one-time O(EnkF)O(EnkF) training cost (epochs EE, noise multiplicity kk, forward cost FF) plus per-example clipping. Our sampler trains nothing. Its private block (subsampling, kernel weighting, clipping and the Gaussian step) costs O(KBd)O(KBd) per output. For a finite release for our experiments at B≤200B \leq 200, our total cost is comparable to ordinary public sampling. Empirically, a 50,000-image CIFAR-10 release takes about 2.5 L40S GPU-hours.

Clarity

Section 2.2 now gives GDP's hypothesis testing interpretation, Section 2.3 opens with PLD's role in obtaining certified bounds, and Section 4 defines all quantities before Proposition 4.1 and explains the proof in three steps: compose within the block, amplify once, compose over outputs.

We would appreciate your assessment of whether these changes resolve your two main concerns.

Reviewer_gMB8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks how much privacy leaks through the diffusion sampling process itself, as opposed to training. Since nobody knows how to bound the sensitivity of a neural denoiser, the authors swap in an empirical denoiser. They clip each per-example term x₀·p(x_t|x₀) to norm C and replace the data-dependent normalizer with a public constant β, which gives replace-one sensitivity 2C/|D|. The Gaussian noise the sampler already adds then serves as the noise of a Gaussian mechanism, so each step is μ_t-GDP and full-batch sampling composes exactly. Since full batch is far too expensive (ε ≈ 200 for one recognizable sample), they propose output-wise subsampling: draw one minibatch of size B per output, reuse it for the whole private block, and apply amplification once per output (Prop. 4.1). All ε are reported as pessimistic PLD bounds from Google’s dp_accounting. Edgeworth is kept only as a diagnostic, and they show it understates ε once ε gets past roughly 5.

Critical strengths
  • The authors are careful to report certified bounds, not approximations. They report the pessimistic PLD bound and show where Edgeworth and the CLT go wrong, by 10–15% and by more than 200× respectively. That’s honest, and useful to anyone doing f-DP accounting with very heterogeneous per-step μ's.
  • The output-wise subsampling design and its accounting are correct. Reusing the minibatch keeps trajectories coherent, and they don’t make the mistake of claiming per-step amplification for it. This is the most reusable idea in the paper.
Critical weaknesses
  • I'm not convinced the private block helps. The obvious baseline is missing: what FID does the public model get alone, with ε = 0? Several things in the paper make me suspect it would be close to the reported numbers:
    • In Table 8, FID gets better as B (and q) shrinks. The authors' own explanation is that a smaller B "weakens" the empirical denoiser.
    • With 9 re-noising steps, FID is around 30–31 no matter where the handoff is.
    • The really good FIDs (e.g., 16.4) are at ε ≈ 108, and even the headline ε ≈ 11 gives a near-trivial ASR bound by the paper's own Appendix L.
    • At ε below 1, FID sits around 30–46, and without the public-only number I can't tell whether that's good.
    • FID is computed against the training set, and a kernel over 25 images will snap to individual training images at low noise. So I'd want to see a nearest-neighbor check before trusting that the late-handoff gains aren't partly copies.
    • Compared with the DP baselines in Table 2, the method is quite a bit worse (27.1 vs. 8–10 on CIFAR), and their ε covers unlimited sampling while this ε covers exactly 50k images.
  • The proxy evidence is weaker than it reads.
    • The cosine in Fig. 1 is between x_t − Ê and x_t − D, and at high noise both are dominated by x_t itself. A value near 1 there is close to guaranteed and says little about the denoisers.
    • The MSEs in Table 4 (up to ~14.6) suggest the unnormalized, clipped drift is far off in magnitude.
    • Table 1 swaps one window at a time inside an otherwise neural trajectory. The actual sampler runs the empirical denoiser from t = 80 all the way to t₀, and the closest number to that ("all steps") has DINO similarity 0.12.
    • The phrasing that leakage "concentrates in critical windows" also seems off. μ_t is fixed by the noise schedule, increases monotonically as t shrinks, and peaks at the very last steps, regardless of where semantics form.
Questions for authors
  1. What's C, and how often does clipping (or numerical underflow of the Gaussian kernel in 3072 dimensions) kick in along the trajectory? I'm honestly unsure what Ê computes in practice with β = 1.
  2. What's the FID of the public model alone, both from scratch and with the same re-noising schedule?
  3. Can you show nearest-neighbor distances from generated samples to training images vs. to held-out images, especially for the 65–70% handoffs?
  4. Which dp_accounting event and neighboring relation did you use, and is μ̄_B computed with 2C/(βB)? It would help to cite Dong et al.'s subsampling theorem directly in Appendix C, which reads a bit informally right now.
  5. Could you run the substitution experiment on the actual private block (empirical from the start to t₀, then the real hybrid continuation), and use a direction metric that doesn't share x_t, e.g. cos(Ê − x̄, D − x̄)?
  6. Why compare at 11.38 against baselines at 10 instead of picking a configuration with ε_u ≤ 10?
  7. Nice to have but not decisive: identity-level privacy for CelebA (many images per person), and an actual membership-inference audit of your sampler. Table 13 attacks a fully trained neural model, which is a different mechanism.
3 comments

Thank you for the careful review and recognizing the relevance of sampling-time privacy. We agree that a public-only comparison and a test of the actual empirical denoiser are essential.

Does the private block help?

Thank you for your question on the baselines for our hybrid denoiser. We provide the two baselines you requested: Public only samples the public denoiser from noise, Schedule-matched follows our hybrid trajectory exactly (same noise, steps, re-noising, public suffix) but uses the public denoiser in the private block too.

SamplerCIFAR-10 FIDCelebA-64 FID
Public only30.70±0.1730.70 \pm 0.1749.26±0.0949.26 \pm 0.09
Schedule-matched control34.68±0.0434.68 \pm 0.0450.97±0.0950.97 \pm 0.09
Hybrid (ours)24.88±0.1024.88 \pm 0.10 (ε≤6.29\varepsilon \le 6.29)22.03±0.1022.03 \pm 0.10 (ε≤3.89\varepsilon \le 3.89)

FID is the mean ±\pm std over three seeds. The private block improves FID over the schedule-matched control by 9.80 on CIFAR-10 and 28.94 on CelebA-64, and over public-only sampling by 5.82 and 27.23. The matched control isolates the private block's contribution, since the restart schedule alone costs the public model about 4 FID on CIFAR-10. We include these baselines in our headline table.

Handing more of the trajectory to the public denoiser moves FID toward the public-only value. On CIFAR-10, for example, nine re-noising steps at 65% give FID 30.30, and with three steps, the hybrid first beats public-only sampling at 62% (FID 29.66, ε≤1.69\varepsilon \le 1.69).

Budget matching

We agree that Table 2 should report a certified bound at or below the baseline budget. The submission matched the Edgeworth estimate to the baseline budget, but Edgeworth is an estimate, and here understates the certified value (10 vs. 11.38 on CIFAR-10, 5.00 vs. 6.77 on CelebA-64). We replace the row in Table 2 for CIFAR-10 and add the row for CelebA-64. FID is mean ±\pm std over three seeds.

Datasetδ\deltaNew Rows for Table 2
CIFAR-1010−510^{-5}FID 24.88±0.1024.88 \pm 0.10, ε≤6.29\varepsilon \le 6.29
CelebA-6410−610^{-6}FID 22.03±0.1022.03 \pm 0.10, ε≤3.89\varepsilon \le 3.89

Our guarantee has a different scope from the others. DP-LDM's (Liu et al., 2024) ε=10\varepsilon = 10 covers a trained model and thus unlimited samples, while ours covers a 50,000-image release, with no network trained on the private data.

As we show in our response to Reviewer KTtX, an LDM fine-tuned only on the released samples covers unlimited sampling at the same ε\varepsilon (FID 22.64 and FID 18.90 for CIFAR10 and CelebA respectively).

Critical windows

Thank you for catching this. To be clear, critical windows are a utility notion and are separate from where privacy loss occurs. μt\mu_t grows as tt decreases and peaks at the final steps (Figure 7), which is why we hand off before them. Critical windows only tell us where to place the handoff in terms of utility and coarse semantics. We correct all instances of this in our paper.

Nearest neighbours

Thank you for this question. You interpreted the mechanism correctly. At low noise, the kernel can concentrate on a single minibatch record, and the public denoiser then fills in the details. We observe that the near-copy rates grow with ε\varepsilon. We demonstrate that late handoffs partly copy training images while earlier handoffs give more novel samples.

For each configuration, we generated 10,000 samples and, following Carlini et al. (2023), computed the root-mean-square (RMS) pixel distance (0–255) to the nearest training and held-out image. We call a sample a near-copy if it is closer to a training image than 99% of held-out images are, which sets a threshold of 19.0 for CIFAR-10 and 24.0 for CelebA. The table below is an ablation showing how the hybrid configuration and ε\varepsilon affect the near-copy rate on CIFAR-10 and CelebA.

SettingCIFAR-10 ε\varepsilonCIFAR-10 near-copiesCelebA ε\varepsilonCelebA near-copies
Public only00.2%00.0%
62%, B=25B=251.694.6%0.7725.1%
65%, B=25B=256.2912.8%3.8950.4%
67%, B=25B=2525.7047.4%13.6690.6%
70%, B=25B=25108.4492.6%53.0299.6%
65%, B=50B=5011.3816.9%5.5057.9%
65%, B=100B=10022.7822.5%8.7466.4%
65%, B=200B=20049.5830.4%15.7771.9%

The main point of this table is that the near-copy rate rises with ε\varepsilon from increasing the number of empirical denoiser steps and batch size BB, from 4.6% at the 62% handoff (ε=1.69\varepsilon=1.69) to 92.6% at the 70% handoff on CIFAR-10 (ε=108\varepsilon = 108). At equal ε\varepsilon, a later handoff copies more than a larger BB. This is expected, since ε\varepsilon is a worst-case bound and does not predict the realized rate.

This is consistent with the certificate. ε\varepsilon bounds how much any single record can change the distribution of the release. The near-copy rate measures the memorization within a share of samples rather than individual records themselves. At the 65% handoff, at most 2.6% of CIFAR-10 and 3.1% of CelebA training records are near-copied. However, the guarantee does not prevent all coarse image structure similarity, especially for larger ϵ\epsilon. We state this phenomenon explicitly in our manuscript.

Held-out FID is similar to the training FID (CIFAR-10: 29.76 vs. 27.44 for ours, 35.93 vs. 33.49 for public only). DINOv2 neighbours detects at most 0.04% of samples are near-copies and find the true source only 2.5% of the time, so we measure in pixel space for better interpretability.

Comparison with FID. We would like to note that the monotonicity intuition (higher privacy cost must produce better image quality) does not necessarily apply to the hybrid sampler. In our setting, smaller batch sizes also improve FID for hybrid denoiser, likely because smaller qq weakens the high-variance empirical denoiser before the public neural denoiser recovers global structure after re-noising.

We include the nearest-neighbour experiment to the revised manuscript and add closest-match figures ranked by the ratio of training to held-out distance.

  • Carlini et al. Extracting Training Data from Diffusion Models. USENIX Security 2023.

Denoiser agreement

Your criticism of Figure 1 is fair. At high noise, the shared term xtx_t can dominate both residuals and give a high cosine similarity even when the denoiser outputs differ. We now describe Figure 1 only as agreement of update directions and no longer use it as evidence that the denoiser outputs agree. Our new figure compares the denoisers using the cosine similarity between E^\hat E and DD and the absolute and relative discrepency gap ∥E^−D∥\|\hat E - D\|. We also measured a cosine centred by the training mean, in DINOv2 and Inception feature space. Both show that the denoisers agree at high noise and separate in the mid-noise window.

Prefix swap. As you suggested, we run the neural sampler from t=80t = 80 to t=t0t = t_0, replace its denoiser with the empirical denoiser for an initial share of the steps (prefix), and let the neural denoiser finish the trajectory (suffix). The initial noise and all sampler noise are shared, so the two differ only in the prefix denoiser. We report the CLIP and DINO cosine similarity between each swapped sample and the neural-only sample from the same noise. (t0t_0 is the noise level handed to the neural denoiser; CIFAR-10, N=39N = 39, 2,048 pairs).

Prefix20%30%40%43%45%50%62%70%100%
t0t_022.811.14.933.973.171.980.520.200
CLIP0.9890.9890.9600.9390.9160.8780.8470.8410.833
DINO0.9930.9930.9150.7800.5650.2770.1730.1520.132

CLIP and DINO measure coarse semantics and instance-level structure, respectively. Both metrics stay above 0.9 up to 40% of generation. DINO then drops sharply, to 0.28 at 50% and 0.17 at 62%. CLIP falls more slowly and levels off near 0.84. Similar to our window-swap experiment, past the mid-noise range, the empirical denoiser keeps the coarse content but not instance-level structure. Past 62% generation, the denoiser swap gives CLIP 0.84 to 0.85, while re-drawing only the prefix's sampler noise gives 0.92 to 0.93. In this range, swapping the denoiser changes the sample more than the sampler’s own randomness does, so it is likely the two denoisers diverge.

We therefore frame the empirical denoiser as a reasonable proxy to the neural denoiser up to about 40% of generation, similar in range to our window-swap experiment. The similarity between the empirical and neural denoiser does not affect the privacy guarantee of our samplers.

We include this experiment alongside the window-swap trajectory in our revised manuscript.

Subsampling and the accountant

We use add/remove adjacency. Each output draws one minibatch of fixed size BB without replacement and reuses it for the private block, so q=B/nq = B/n. An added record displaces another one, so the minibatch average has sensitivity 2C/B2C/B, μt=2CβB2Δt/t3\mu_t = \frac{2C}{\beta B}\sqrt{2\Delta t/t^3}, and μˉ=(∑tμt2)1/2\bar\mu = (\sum_t \mu_t^2)^{1/2} over the private block. Appendix C now cites the subsampling theorem of Dong et al. (2022) directly, and C.1 adapts it to add/remove neighbors with a coupling (Zhu et al., 2022; Lebeda et al., 2025). In Google's PLD Accounting Library (Doroshenko et al., 2022) we call from_gaussian_mechanism(1/mu_bar, sampling_prob=q_B) with pessimistic discretization (h=10−5h = 10^{-5}) under ADD_OR_REMOVE_ONE, then self_compose(N_out). Its two directions are exactly hBh_{\mathcal B} and hB−1h_{\mathcal B}^{-1} in Proposition 4.1. The reported budgets are unchanged.

  • Doroshenko et al. Connect the Dots: Tighter Discrete Approximations of Privacy Loss Distributions. PETS 2022.
  • Lebeda et al. Avoiding Pitfalls for Privacy Accounting of Subsampled Mechanisms under Composition. SaTML 2025.
  • Zhu et al. Optimal Accounting of Differential Privacy via Characteristic Function. AISTATS 2022.

What does the empirical denoiser compute?

We use C=50C = 50. At native 32x32 resolution, the median is 26.9, and 0.02% of records exceed CC. For 64x64, record norms have median 53.5 (5th–95th percentile 33.1–78.8). The kernel exp⁡(−∥xt−yi∥2/2t2)\exp(-\|x_t - y_i\|^2/2t^2) is computed in log space with the log-sum-exp stabilization to prevent the underflowing of the Gaussian kernel. With the log-kernel ℓi\ell_i, the weights are wi=exp⁡(ℓi−s)∈(0,1]w_i = \exp(\ell_i - s) \in (0, 1], and with β=1\beta = 1 and s≈max⁡j∈Sℓjs \approx \max_{j \in S} \ell_j, the empirical denoiser is E^=∑i∈Sclip(wiyi,C)\hat{E} = \sum_{i \in S} \mathrm{clip}(w_i y_i, C). Without clipping, this is the minibatch posterior mean scaled by ∑iwi∈[1,B]\sum_i w_i \in [1, B]. Since wi≤1w_i \le 1, a contribution can only be clipped if its record's norm exceeds CC. The nearest record always has weight 1, so it is clipped whenever its norm exceeds CC.

Future Work

We leave identity-level (group) privacy for CelebA and a formal membership-inference audit of the sampler to future work, and note that Table 14 reports an attack on a trained model rather than on our sampler.

Reviewer_C9WrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper investigates introducing the DP gausssian mechanism during the sampling phase of the diffusion generators (mostly data plane in their experiments, i.e, pixel space), the The question that the papers is asking is intersting and important. But the scope of experiments and a definition of a real threat model for this paper is missing, among some unjustified and unintuitive claims.

Strengths and Weaknesses

Generally the paper is well-written the writting on supported claims can be imporoved, and I suggest the authors to not use indirect pharases, like through out the text you find a lot of, this will lead to x equals to y, not z .

L38-L41 → Is there a study that shows this or have the authors considered showing this first? e.g., performing an attack on models that applied post-hoc that stills shows their training data?

Since the sampler is a tool that is usually outside the reach of the models contributor, how is the threat model and attack model is being managed in this case? It would be nice to put a paragraph on clarifying how this method incorporates in real use case (e.g., released images in this case I believe).

There are tones of ways that researchers do time-step distilliation to few step samplers (and of course many other methods) does the authors investigated how these type of models will evaluated against this mechanism that introduced for multi-step procedures?

L249-251: Does this statement holds for conditional versoins of the generator too? e.g., conditional, cifar10 or imagenet? it was not clear for me if the authors used edm or edm2 models, in eitherway there is a conditional version of the generator too.

Its not clear the cosine is in what space, Fig.1, is it in the x space or a feature space? how meaningful this metric can be? is everything in pixel space, an investigation of a this mse or cosine in the latent space might be interesting to do?

Clarification of the Table.1 might be necessary, what is the key result that the authors want to show to the reader? e.g., they provide a privacy loss dist with the equivalent similar epsilons against other methods? maybe its better to introduce another column that clearly states this that e.g., with method X we cannot have a certified PLD.

Minor

Minor: L125: iff

L128: maybe rewording required? is a functional of it

How does the ranges are depicted in Table1? t range

Critical strengths

Please see the paper summary

Critical weaknesses

Please see the paper summary

Questions for authors

Please see the paper summary

2 comments

Thank you for recognizing the importance of private sampling, and for your suggestions on the threat model and the experiments. We address the concerns about the threat model, few-step sampling, and conditional generation.

1. Post-hoc defenses (L38–41)

Thank you for the question. Yes, there are similar studies that exist for text-to-image diffusion models and language models. The closest work to what you describe is Kowalczuk et al. (2026). They show that small perturbations of the text embedding make the model reproduce its training images, even for models with memorization defense, such as NeMo (Hintersdorf et al., 2024). For language models, Ippolito et al. (2023) build a filter that blocks every verbatim training sequence and still extract training data with modified prompts. Both defenses act through text conditioning, while our models are unconditional. More importantly, none of these attacks can settle our research question, because their attack's success does not provide a privacy guarantee.

We revise that line to state that post-hoc filtering and memorization mitigation do not by themselves give a DP guarantee, and that stronger attacks have bypassed several of them, citing these papers.

  • Hintersdorf et al. Finding NeMo: Localizing Neurons Responsible for Memorization in Diffusion Models. NeurIPS 2024.
  • Ippolito et al. Preventing Generation of Verbatim Memorization in Language Models Gives a False Sense of Privacy. INLG 2023.
  • Kowalczuk et al. Finding DoRI: Discovery of Retained Images in Diffusion Models. ICML 2026.

2. Threat model and practical use

A trusted data holder releases a synthetic dataset, or runs a generation service that releases a number of outputs at a privacy budget. For each output, it draws a hidden uniform minibatch, runs the clipped empirical private block, and completes generation with a public model trained independently of the private data. The guarantee covers over the entire released samples. This requires control of the private sampler and of the release count. Once the samples are released, the images can be used freely, and any model trained only on them and public data inherits the guarantee by post-processing. In our response to KTtX, we fine-tune a latent diffusion model on the released samples, and its samples carry the same guarantee with no new accounting.

An adversary can have access to all released samples and may know the algorithm and the public model. What stays private is the dataset itself and the sampler's randomness.

We updated the manuscript to reflect these changes.

3. Few-step and distilled sampling

Thank you for this interesting suggestion. We have not evaluated on distilled samplers, but we can extend our analysis to them. The accounting depends only on the sensitivity and the noise of each private step, so the key question is which denoiser touches the private data. The analysis covers any private step of the form

xi+1=aixi+bifD(xi)+siz,z∼N(0,I),x_{i+1} = a_i x_i + b_i f_D(x_i) + s_i z, \qquad z \sim \mathcal{N}(0, I),

where aia_i, bib_i and si>0s_i > 0 are public and ∥fD(x)−fD′(x)∥≤Δi\|f_D(x) - f_{D'}(x)\| \le \Delta_i for every state xx and neighboring D,D′D, D'. Each such step is a Gaussian mechanism with μi=∣bi∣Δi/si\mu_i = |b_i|\Delta_i / s_i. In our sampler, fDf_D is the clipped empirical denoiser, and the EDM step (ai=1−2Δt/ta_i = 1 - 2\Delta t/t, bi=2Δt/tb_i = 2\Delta t/t, si=2tΔts_i = \sqrt{2t\Delta t}, Δi=2C/(βB)\Delta_i = 2C/(\beta B)) recovers μt\mu_t from Section 4. A few-step schedule only changes aia_i, bib_i and sis_i, so its guarantee follows by recomputing μi\mu_i and composing as in Section 4. We find the following:

  • Fewer steps do not by themselves lower the privacy loss, because a larger step also raises the drift-to-noise ratio ∣bi∣/si|b_i|/s_i.
  • Deterministic private steps (si=0s_i = 0, as in DDIM with η=0\eta = 0; Song et al., 2021) add no noise and fall outside the analysis.
  • A distilled network trained on the private data and used inside the private block has no tractable sensitivity bound, the same open problem as for any neural denoiser (Section 3.3).
  • A distilled model trained only on public data and used after the handoff is post-processing, so it can make generation faster with the same guarantee.

We list these conditions in our limitations and future work.

  • Song et al. Denoising Diffusion Implicit Models. ICLR 2021.

4. Conditional generation and model details (L249–251)

All samples in the paper are generated unconditionally, without class labels or guidance. The only neural networks in our sampler are the public denoisers used after the handoff. The empirical denoiser is computed directly from the private records and has no architecture or weights. For CIFAR-10, the public denoiser is a class-conditional EDM model (Karras et al., 2022) trained on ImageNet 32x32, which we sample without labels by passing a null label. For CelebA-64, it is an unconditional EDM model trained on FFHQ 64x64. Figure 1 and Table 1 use a separate unconditional EDM model trained on CIFAR-10 32x32 as the neural reference, because the proxy question is how closely the empirical denoiser follows a network trained on the same data. We describe the model details in Section 6.

We thank the reviewer for pointing out the case for conditional generation. We did not run conditional experiments, but our analysis can extend directly to one form. Suppose each output's condition, such as a class label, is chosen without looking at the private data. A class-conditional empirical denoiser can then keep the same uniform subsample and the normalization βB\beta B, and give zero weight to records from other classes. Clipping still bounds each record's contribution by CC, so the per-step sensitivity stays at most 2C/(βB)2C/(\beta B) and the accounting is unchanged. However, conditioning that depends on the private data is a different mechanism that needs its own accounting. We add this case as future work.

  • Karras et al. Elucidating the Design Space of Diffusion-Based Generative Models. NeurIPS 2022.

5. Figure 1 and the measurement space.

Figure 1 reports the pixel-space cosine similarity of the residuals, cos⁡(xt−E^,xt−D)\cos(x_t - \hat{E}, x_t - D). We chose it because these residuals choose the direction of each denoising update. The paper also reports the MSE between the denoiser outputs (Table 4), CLIP and DINO features of the final samples (Tables 1 and 5), and pixel-space PSNR, SSIM and label agreement (Table 6). Following your suggestion, we repeated the comparison in feature space with DINOv2 CLS embeddings (Oquab et al., 2024), centred by the training-set mean, and comparing with held-out test images. We include the figure in our revised version. The main experiments are in pixel space, but a latent-space version is interesting future work.

6. Table 2 and its ranges.

Table 1 does not report privacy values, so we believe the question refers to Table 2. Nonetheless, we will clarify both tables.

In Table 1, tt denotes the noise standard deviation, and each range identifies a selected window of the denoising trajectory. The goal is to compare the final output of a neural-only trajectory against the final output of the trajectory with the same noise, but swapping the empirical denoiser for some stretch of time (which we call the window). "swap" measures the final CLIP/DINO cosine distance after replacing the neural denoiser with the empirical denoiser within that window. "reference" acts as a control, measuring the final CLIP/DINO cosine distance after re-noising the neural trajectory. Their ratio provides a comparison and determines whether the empirical denoiser window results in a similar final output as with the neural. We use this to show when empirical denoisers could be a reasonable proxy for a neural denoiser. We understand that the metrics may be confusing and we swap them with the more intuitive CLIP/DINO cosine similarity.

For Table 2, our methods shows the certified ε\varepsilon for the joint release of Nout=50,000N_{\rm out} = 50{,}000 samples. We use PLD accounting and report both the certified upper bound and lower bound from discretization slack. The other methods in Table 2 are mostly based on DP training, so their reported ε\varepsilon certifies a trained model.

We provide a table in the revision that demonstrates what each method certified.

7. Clarity

We expanded "iff" to "if and only if", replaced "is a functional of it" with an explicit statement of how the PLD determines δ(ϵ)\delta(\epsilon) (Section 2.3), and revised indirect phrasing throughout.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.