acceptodds
Under review as a conference paper at ICLR 2027

Protecting Sensitive Data in Image Synthesis via PAC-Private Adaptation for Diffusion Models

Abstract

Synthetic data are increasingly used as an alternative to sharing sensitive records. However, synthetic data generation does not guarantee privacy, as diffusion models trained or adapted on sensitive data remain susceptible to reconstruction attacks. Moreover, while approaches that use differential privacy (DP), such as DP-SGD, achieve provably private diffusion model training, the repeated gradient clipping and noise injection they require result in significant utility loss. An important limitation of DP-based privacy is that, although it has a provable relationship to reconstruction privacy (RP), that relationship is indirect. RP is defined in terms of limiting how much an adversary’s posterior distribution over sensitive data differs from the prior, whereas DP provides guarantees by bounding the sensitivity of outputs to changes in individual records. This indirection is an important source of the utility loss. To address this, we propose a PAC-private diffusion model adaptation to achieve reconstruction privacy. Since PAC-privacy is defined directly with respect to posterior advantage over the prior, it directly implicates RP. To obtain scalable PAC privatization in high dimensions, we first learn a compact data-dependent diffusion model component using LoRA or Textual Inversion, and then calibrate anisotropic Gaussian noise from the covariance of repeated mechanism outputs. Unlike DP-SGD, our method perturbs the learned component only once after optimization, thereby avoiding privacy composition across gradient updates. We evaluate the framework on few-shot concept personalization and full-dataset image synthesis, and show that the proposed approach better preserves subject identity, generation quality, and downstream classification accuracy than DP while achieving the same reconstruction privacy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Reconstruction privacy, Diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7P8kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a PAC-private diffusion model adaptation to directly achieve reconstruction privacy (RP) to overcome the limitation of DP-SGD that it is does not have a direct relationship to RP. A diffusion model private adaptation is done with PAC-privacy for sensitive image synthesis. It trains compact LoRA adapters or Textual Inversion embeddings and applies a single Gaussian perturbation after training, with anisotropic noise calibrated from the covariance. It also canonicalizes LoRA factors to reduce parameterization ambiguity. It tried to match the privacy budge between DP with PAC privacy which is later used for experimental comparisons. The reported results under their matching mechanism show improved subject fidelity and classification accuracy over DP-SGD.

Critical strengths
  1. The motivation is clear and promising. The observation that DP is not directly related to how much information is revealed before and after private training is a long-standing vital problem of the field of privacy. I am happy to see progress in this field.
  2. The experiments cover both personalization and downstream synthetic-data utility, which is really nice since it does not stays only on the "theoretical" part, but moves toward the "application" part.
Critical weaknesses
  1. My biggest concern is about the theoretical part. I did not reproduce the proof myself, but the claimed privacy-budget matching between PAC privacy and DP is insufficiently justified. The conversion matches reconstruction-success bounds only at the selected prior success probability α⋆=(1−δ)/(1+eε)\alpha^\star=(1-\delta)/(1+e^\varepsilon). There are 22 concerns to me: (1)(1) the tightness of these two upper bounds are not discussed, so they may or may not give the same constraint even when the value is equal; (2)(2) even when we assume that they exactly equal, one α⋆\alpha^\star is matched to millions of different pairs of (ε,δ)(\varepsilon,\delta) which forms a continuous "line". Consequently, I doubt that the utility comparisons is not conducted under equivalent reconstruction privacy.
  2. Another concern is about the theoretical proof. I did not reproduce the theorems and proofs myself, but I used GPT-6.1-Sol (extra-high) to give them a check. The LLM pointed out a major concern that I checked myself and believe to be real. On page 5, in paragraph Noise calibration., ΣB\Sigma_{B} is drawn from private training dataset XX, and then noise is drawn from it in Theorem 1. This contradicts with the assumption also in Theorem 1 that the noise is independent of X⋆X^\star, the private dataset used for the final adaptation. If the same dataset is used in the experiment, I would completely question the soundness of the results. Correct me if I make any mistakes or misunderstandings here.
  3. Although PAC privacy is the key part of the paper, the formulation of it is never given throughout the paper, even with the formulation of DP given.
  4. I am bit concerned that the tasks used in the experiment part is too simple.
Questions for authors

See Weaknesses. Especially the theoretical part, which is hard be really vital given the topic and scope claimed by the paper.

3 comments
AuthorsOpenReview ↗

Author Response

Thank you for recognizing the importance of studying reconstruction privacy and for appreciating our evaluation of both personalized generation and downstream synthetic-data utility. We also appreciate your close reading of the theoretical analysis and the questions it raised. Below, we address the budget conversion, noise calibration, PAC Privacy formulation, and experimental setting in turn.

Q1: Is the DP-to-PAC budget conversion justified when the reconstruction success bounds are matched at only one prior success rate?

Thank you for raising the question of non-uniqueness. The trade-off function fε,δf_{\varepsilon,\delta} is an equivalent representation of the (ε,δ)(\varepsilon,\delta)-DP guarantee and thus provides a valid posterior reconstruction-success bound at each prior success rate. These are analytically derived upper bounds, not measurements of realized attack success. Privacy–utility comparisons ask which method achieves higher utility under the same stated privacy constraint. Here, we use the upper bound on posterior reconstruction success at α⋆\alpha^\star as that constraint: we choose the PAC mutual information budget to match the DP bound at this point, then compare utility of DP-SGD and PAC perturbation.

We agree that picking different points on the f-DP curve yields different values of KL divergence (β\beta). Conversely, a fixed β\beta is compatible with many prior–posterior success pairs and may be associated with different DP parameters as well.

Prior success rate α\alphaβ\beta at ε=3\varepsilon=3β\beta at ε=5\varepsilon=5β\beta at ε=10\varepsilon=10
10−610^{-6}0.0000730.0006450.198584
10−510^{-5}0.0004420.0059982.008758
4.5397×10−54.5397\times10^{-5} (α10⋆\alpha^\star_{10})0.0018990.0270689.999102
10−410^{-4}0.0041490.0596259.209423
10−310^{-3}0.0413850.6061596.906943
6.6928×10−36.6928\times10^{-3} (α5⋆\alpha^\star_{5})0.2841704.9330825.006004
10−210^{-2}0.4314514.5344514.604469
4.7425×10−24.7425\times10^{-2} (α3⋆\alpha^\star_{3})2.7154572.9905413.047990
10−110^{-1}2.0211982.2522562.302042
0.50.50.5766320.6706040.692882

Choosing a prior success rate arbitrarily gives a valid pointwise comparison, but may obscure differences in the strength of DP guarantees. For example, at α=0.1\alpha=0.1, the resulting KL budgets are 2.021, 2.252, and 2.302 for ε=3,5,10\varepsilon=3,5,10, respectively, despite the substantial differences in ε\varepsilon. We therefore do not choose the prior success rate arbitrarily. Instead, we choose the prior success rate where DP permits its maximum reconstruction advantage, and compute the corresponding KL budget. With δ\delta fixed, β\beta is strictly increasing in ε\varepsilon, giving a unique mapping under this selection rule. Thus, the conversion matches the reconstruction-success upper bounds at the chosen operating points and preserves the ordering of the DP budgets as ε\varepsilon increases.

Q2: Is ΣB\Sigma_B independent of X⋆X^\star? Thank you for raising the independence question. Let DD be the fixed dataset, and let Q(D)\mathcal Q(D) denote the distribution induced by our resampling rule. Calibration uses independent draws X(1),…,X(K)∼Q(D)X^{(1)},\ldots,X^{(K)}\sim\mathcal Q(D), while the final adaptation uses a fresh draw X⋆∼Q(D)X^\star\sim\mathcal Q(D), independent of the calibration draws conditional on DD. The estimated noise covariance depends only on the calibration runs, and the Gaussian noise is drawn afresh. Consequently, the added noise is independent of X⋆X^\star. Sampling from the same fixed distribution does not make the calibration and final draws dependent.

Q3: PAC Privacy definition.

Due to space constraints and readability, we presented the reconstruction advantage and mutual information formulation directly used in our analysis, while citing the original PAC Privacy paper for the full definition.

Q4: Experimental tasks.

Our evaluation builds on established private image-generation tasks: DPDM evaluates FashionMNIST and downstream classification using synthetic data, while DP-LoRA studies private latent-diffusion fine-tuning. We additionally evaluate EuroSAT, BloodMNIST, and few-shot object and face personalization to test the methods across different data domains and adaptation settings.

Thank you again for your careful review. We hope our responses clarify and address your concerns.

Thank you for the clarification. I agree that independent resamples from a fixed dataset can be independent. However, this does not resolve the central concern: the response changes the privacy model without establishing the guarantee claimed in the paper. The manuscript explicitly aims to protect the original sensitive training records (lines 151–165), formalizes privacy through I(X;M(X))I(X;M(X)) (lines 180–182), and claims I(X⋆;ϕ~)≤βI(X^\star;\widetilde{\phi})\leq\beta for X⋆∼DnX^\star\sim\mathcal D^n in Theorem 1 (lines 262–269). Algorithm 1 instead calibrates noise using bootstrap samples from the private dataset XtrainX_{\mathrm{train}} (lines 220–228 and 233–235). Your explanation establishes independence conditional on XtrainX_{\mathrm{train}}. This can justify independence within a fixed-source resampling model, but it does not establish confidentiality of the source records themselves. If XtrainX_{\mathrm{train}} is the private object to be protected, drawing fresh Gaussian randomness does not remove the dependence of the noise distribution on that dataset. Please explicitly reconcile the protected variable, sampling distribution, and adversary’s prior across the privacy definition, theorem, and experimental procedure.

Furthermore, I agree with the first point raised in "Critical Weaknesses" by Reviewer 3PUr. Therefore, I believe the theoretical part of this paper needs deeper revisits. I recommend the authors use some cutting models for help in re-check and re-verification of the proposed theoretical part.

AuthorsOpenReview ↗

Response to the Follow-up

Thank you for the follow-up. We agree that the distinction between the protected secrets and their underlying probability distribution should be made explicit.

In our PAC Privacy formulation, the privacy guarantee concerns information leakage about the realized sensitive data, modeled as random variables under a specified input distribution, rather than the distribution itself. In particular, let X⋆∼DnX^\star\sim\mathcal D^n denote the dataset used for the final adaptation, where D\mathcal D specifies the underlying data distribution. The guarantee bounds the mutual information between X⋆X^\star and the released output, thereby covering the sensitive records contained in X⋆X^\star. This is a distributional privacy guarantee and is not tied to any particular realization of X⋆X^\star.

Our Monte Carlo calibration assumes access to a sampling procedure that generates calibration datasets independently from the same fixed input distribution Dn\mathcal D^n. Crucially, the calibration procedure does not access the realized protected data in X⋆X^\star or any information derived from them. Since the calibration datasets and X⋆X^\star are sampled independently, the estimated noise covariance and the Gaussian perturbation generated from it are independent of X⋆X^\star. This establishes the independence condition underlying Theorem 1.

We will revise the algorithm description and privacy analysis to explicitly distinguish the realized protected secrets, their underlying probability distribution, and the independent calibration procedure. When an empirical distribution is constructed from a reference dataset, the independence argument requires that the reference information used for calibration be independent of the protected input. The stated guarantee does not automatically extend to records used to construct that empirical distribution.

Thank you also for considering Reviewer 3PUr's comments. We have addressed those concerns in our separate response.

Reviewer_3PUrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes using PAC privacy instead of DP-SGD to privatize diffusion model adaptation. The idea is to train a small adapter (rank-1 LoRA on the cross-attention K/V projections, or a Textual Inversion embedding) many times on bootstrap resamples of the private data. The authors then estimate the covariance of the resulting adapters and add anisotropic Gaussian noise once to a final trained adapter, using the noise allocation rule from Sridhar et al. (2025). Theorem 1 bounds the mutual information between the data and the released adapter by β, and through the Xiao & Devadas result this gives a bound on reconstruction advantage ("β-RP"). For LoRA they also canonicalize the factors with an SVD before estimating the covariance, to remove the scale/sign ambiguity. To compare against DP, they convert (ε, δ) into a β using the f-DP reconstruction bound of Kulynych et al., evaluated at one particular prior success rate.

Critical strengths

The core design makes sense to me. Privatizing a small adapter once, rather than clipping and noising every gradient step, removes the composition overhead that hurts DP-SGD so much with tiny datasets. If the accounting were correct, the reported gains (e.g., DINO-I 0.352 vs 0.149 at ε = 10 on DreamBooth, and 26.4% vs 14.1% TSTR on BloodMNIST) would be practically meaningful.

The experiments are also well documented. Prompts, metrics, hyperparameters, CIs and the release protocol are all spelled out, which made it possible to pin down the concerns below concretely.

Critical weaknesses

The guarantee doesn’t seem to apply to the actual procedure. Theorem 1 assumes the true covariance of the adapter, but the algorithm uses an estimate from K = 96 runs. By my rough count the LoRA adapter has on the order of 50k parameters, and the TI embedding has 768. Either way the estimated covariance has rank at most 95. Under the allocation rule, zero eigenvalues get zero noise, so almost every direction of the released adapter is left unperturbed. The true variance in those directions isn’t zero, so the MI bound breaks. I couldn’t find any regularization, eigenvalue floor or finite-sample correction in the paper.

Two related problems make this worse. First, the covariance comes from bootstrapping the realized private set, not from sampling datasets from D as the definition requires. With only 4–6 DreamBooth images, there are just a few hundred distinct resamples, and I don’t see why their spread would reflect the variance under the adversary’s prior. Second, the noise covariance itself depends on the private data, so the Markov chain argument in the proof (with B independent of X*) isn’t quite right as written. Since there’s also no empirical attack, I can’t tell how private the released adapters really are.

The privacy matching isn’t apples to apples. DP bounds per-record leakage against a strong, informed adversary on worst-case data. β-RP bounds average-case MI about the whole dataset under a prior. The conversion is also done at a single prior success rate, where β works out to roughly ε nats (about 2.7 at ε = 3 and about 10 at ε = 10). Ten nats of information about a five-image dataset is a very weak guarantee. Some of PAC’s advantage may simply come from the weaker notion. The CelebA experiment makes this concrete: PAC “wins” because its identity similarity rises from 0.004 to 0.195, meaning it reproduces the private person’s face much better. That seems to be exactly what Section 3 says the method should prevent.

Baselines. DP-SGD looks under-tuned. It uses clip norm 0.01, the sampling rate is near 1, and its numbers barely move from ε = 3 to ε = 50. I’d also like to see the public model with prompts and no adapter, the clean adapter in the tables, DP output perturbation on a clipped final adapter, and isotropic PAC noise. Without the zero-shot numbers, I can’t tell whether DP-SGD, or PAC at small ε, learns anything from the private data. The cost of 96 full training runs should also be reported.

Questions for authors
  1. How do you handle the rank deficiency of the estimated covariance? Is any noise added outside its span? If not, how does Theorem 1 apply? Please report the adapter dimension for each setting.
  2. Why should bootstrapping one realized dataset approximate the covariance under draws from D? What is D in the DreamBooth and CelebA settings? How do you account for leakage through the data-dependent noise covariance?
  3. What do the zero-shot public model and the clean adapter score on each metric?
  4. Was DP-SGD tuned per ε (clip norm, learning rate, steps)? Why is it so flat across budgets?
  5. How much does the matched β change with the choice of prior success rate?
  6. Could you run an actual reconstruction or identity-matching attack on released adapters for both methods?
4 comments
AuthorsOpenReview ↗

Author Response (1/4)

Thank you for your careful and technically detailed review. We appreciate your recognition of the motivation for perturbing a compact adapter only once, as well as the detailed documentation of our experiments. We first clarify our privacy definition and experimental setting, then address your specific questions with additional baselines, tuning results, and reconstruction-attack evidence.

Clarification 1: The distribution induced by resampling a finite dataset.

For a fixed dataset DD, a resampling rule Q\mathcal Q defines a distribution Q(D)\mathcal Q(D) over possible inputs to the adaptation procedure. Our description was written with large datasets in mind and did not make the few-shot case explicit. In the few-shot experiments, each run randomly selects a subset of the concept's images from DD. We have revised it in the updated version.

Clarification 2: BB is independent of X⋆X^\star.

Calibration and the final adaptation use the same resampling rule, with independent draws. Thus, the calibration outputs ϕ(k)=M(X(k);ξ(k))\phi^{(k)}=M(X^{(k)};\xi^{(k)}), where X(k)∼Q(D)X^{(k)}\sim\mathcal Q(D) and ξ(k)\xi^{(k)} is fresh training randomness, sample the output distribution induced by this rule. Their empirical covariance estimates Cov⁡X∼Q(D), ξ[M(X;ξ)]\operatorname{Cov}_{X\sim\mathcal Q(D),\,\xi}[M(X;\xi)]. The small number of available images limits the support of Q(D)\mathcal Q(D), but does not prevent us from defining or sampling from this distribution. We will clarify that this is the distribution targeted by our calibration, without claiming it represents an unknown broader population of images.

Clarification 3: A dataset-level mutual-information bound also applies to each sampled record.

We write X=(X1,…,Xn)X=(X_1,\ldots,X_n) for the random input dataset because the adaptation mechanism operates on the dataset as a whole. For each ii, XiX_i is a component of XX, so Xi→X→ϕ~X_i\rightarrow X\rightarrow\widetilde\phi forms a Markov chain. By the data-processing inequality, I(Xi;ϕ~)≤I(X;ϕ~)≤βI(X_i;\widetilde\phi)\leq I(X;\widetilde\phi)\leq\beta. Thus, the bound on the sampled dataset also bounds leakage about each sampled record under the specified input distribution. Although this is not a worst-case guarantee over neighboring datasets like DP, it still bounds information leakage about each individual datapoint under the specified input distribution.

Clarification 4: Interpreting the mutual-information privacy guarantee

We agree that β≈10\beta\approx10 nats should not be presented as a strong absolute privacy guarantee. Likewise, ε=10\varepsilon=10, and especially larger values, provides a weak DP guarantee. The PAC budget is derived from these DP parameters by matching the analytical upper bounds on posterior reconstruction success. It is not chosen independently to give PAC more room to preserve utility. Our comparison asks which method retains more utility under this specified pointwise privacy constraint. In our experiments, DP-SGD improves only modestly even as ε\varepsilon increases, whereas PAC retains substantially more utility.

Clarification 5: Image Reconstruction Risk and Identity Fidelity Are Distinct

We appreciate why the higher ISM on CelebA-HQ raises a privacy concern. If the protected secret were the subject’s identity, generating a recognizable face would indeed be relevant leakage. In our few-shot experiments, however, identity fidelity is the intended utility: each adapter is trained to generate new images of one subject. The privacy concern is whether it reproduces a particular photograph used for training. ISM measures identity fidelity, not copying of a training photograph. Therefore, PAC’s higher ISM does not by itself establish or rule out image reconstruction. We will make this distinction explicit in the paper and assess reconstruction directly in our response to Question 6.

Clarification 6: DP Output Perturbation Does Not Close the Utility Gap

DP output perturbation does not consistently outperform DP-SGD and remains well below PAC in subject fidelity across the tested privacy budgets. We randomly selected four DreamBooth subjects for this comparison. For each ε\varepsilon, we selected the clip norm CC with the highest DINO-I. We then evaluated the selected setting on the four subjects and averaged the results over subjects and training seeds.

ε\varepsilonDP output (best development CC)DP-SGDPAC
3.154 (C=0.5C=0.5).150.240
5.155 (C=0.5C=0.5).150.312
10.155 (C=0.5C=0.5).167.352
20.156 (C=0.5C=0.5).180.368
50.181 (C=11C=11).206.375

The frozen-model DINO-I score is 0.190, while the per-image adapter average reaches 0.338 without privacy noise. DP output perturbation therefore remains at or below the frozen baseline despite tuning CC, and it does not close the subject-fidelity gap to PAC.

AuthorsOpenReview ↗

Author Response (2/4)

Clarification 7: Covariance-based anisotropic noise allocation

PAC mechanism allocates Gaussian noise anisotropically according to the covariance of the learning output. This covariance-based allocation is part of the method itself.

Q1: Rank Deficiency and Finite-Sample Calibration

In our experiments, the released component has 49,536 dimensions for rank-one LoRA on DreamBooth and CustomConcept101, 27,264 for rank-one LoRA on BK-SDM-Small, and 768 for Textual Inversion on CelebA-HQ. Theorem 1 gives a valid mutual-information guarantee when the population covariance of that component is known and noise is allocated accordingly.

In practice, following the simulation-based calibration approach of Sridhar et al., we repeat the complete diffusion adaptation procedure until the relative change in the estimated covariance trace falls below a preset threshold. In our experiments, additional draws introduced only small nonzero eigenvalues. We also provide an appendix analysis showing that an explicit full-rank correction to the estimated covariance yields a (1−γ)(1-\gamma)-confidence guarantee under its stated assumptions.

Q2: Why should bootstrapping one realized dataset approximate the covariance under draws from D? What is D in the DreamBooth and CelebA settings? How do you account for leakage through the data-dependent noise covariance?

PAC Privacy measures how much observing a mechanism’s output improves reconstruction success beyond what is possible from the known input distribution alone. In our few-shot setting, the finite image pool UcU_c and the subset-sampling rule define a distribution Qc\mathcal Q_c over possible training inputs. For a nontrivial random subsample, knowing the candidate images does not reveal which ones were selected for the final adaptation. Our privacy objective is to limit the additional ability to identify that selected input after observing the released adapter. The small size of UcU_c does not invalidate this setting. It is reflected in the prior success rate under Qc\mathcal Q_c. We do not claim that UcU_c represents a broader population of unseen images.

The calibration inputs and final input X⋆X^\star are drawn independently from Qc\mathcal Q_c. Conditional on UcU_c, the estimated noise covariance therefore depends on the calibration runs but not on X⋆X^\star: I(X⋆;Σ^B∣Uc)=0I(X^\star;\widehat\Sigma_B\mid U_c)=0. Observing the covariance alone consequently provides no additional information about which images were selected for the final run.

Q3: What do the zero-shot public model and the clean adapter score on each metric?

Thank you for pointing out that the frozen and clean baselines were not sufficiently visible. We have revised the main quantitative tables to report the unadapted pretrained model (frozen) and the non-private adaptation (clean) on every metric for each dataset. For a compact summary, the table below compares their primary metrics with DP-SGD and PAC at ε=10\varepsilon=10.

Setting (primary metric)FrozenCleanDP-SGDPAC
DreamBooth (DINO-I ↑)0.1470.3880.1490.352
CustomConcept101 (DINO-I ↑)0.1260.4050.1410.353
CelebA-HQ (ISM ↑)0.0010.3410.0020.103
FashionMNIST (TSTR accuracy ↑)26.4%60.5%37.2%43.1%
EuroSAT (TSTR accuracy ↑)23.7%47.6%35.5%39.6%
BloodMNIST (TSTR accuracy ↑)10.4%38.1%14.1%26.4%

The detailed scores on other metrics are updated in the revised version of our paper.

AuthorsOpenReview ↗

Author Response (3/4)

Q4: Was DP-SGD tuned per ε\varepsilon (clip norm, learning rate, steps)? Why is it nearly flat across budgets?

To assess sensitivity to clipping, we varied CC on a subset of Dreambooth, holding the other settings fixed. The DINO-I is reported in the table below.

ε\varepsilonPACDP C=.005C=.005DP C=.01C=.01DP C=.03C=.03DP C=.1C=.1
3.137.085.085.084.084
5.218.111.112.105.104
10.293.130.113.112.110
20.316.135.125.146.121
50.333.158.167.185.168

Changing the clip norm of DP-SGD may bring some modest improvements but it is still far below PAC. In fact, as long as the optimization settings are fixed, the utility can reflect the performance of the privatization mechanism. This can be verified using different combinations of the learning rate and optimization steps. Each cell below is DP-SGD / PAC DINO-I.

Learning rate, updatesε=3\varepsilon=3ε=5\varepsilon=5ε=10\varepsilon=10ε=20\varepsilon=20ε=50\varepsilon=50
2×10−42\times10^{-4}, 250 (main).085 / .137.112 / .218.113 / .293.125 / .316.167 / .333
5×10−55\times10^{-5}, 250.082 / .236.107 / .241.122 / .248.147 / .298.169 / .340
1×10−41\times10^{-4}, 250.078 / .201.095 / .226.105 / .234.127 / .242.132 / .250
4×10−44\times10^{-4}, 250.077 / .110.084 / .162.108 / .276.136 / .342.123 / .413
2×10−42\times10^{-4}, 125.087 / .188.092 / .231.101 / .241.131 / .249.133 / .259
2×10−42\times10^{-4}, 500.066 / .120.094 / .180.100 / .202.118 / .297.136 / .368

PAC perturbation achieves higher DINO-I score in all the paired settings.

Q5: What is the trend of β\beta as prior success rate changes?

Prior success rate α\alphaβ\beta at ε=3\varepsilon=3β\beta at ε=5\varepsilon=5β\beta at ε=10\varepsilon=10
10−610^{-6}0.0000730.0006450.198584
10−510^{-5}0.0004420.0059982.008758
4.5397×10−54.5397\times10^{-5} (α10⋆\alpha^\star_{10})0.0018990.0270689.999102
10−410^{-4}0.0041490.0596259.209423
10−310^{-3}0.0413850.6061596.906943
6.6928×10−36.6928\times10^{-3} (α5⋆\alpha^\star_{5})0.2841704.9330825.006004
10−210^{-2}0.4314514.5344514.604469
4.7425×10−24.7425\times10^{-2} (α3⋆\alpha^\star_{3})2.7154572.9905413.047990
10−110^{-1}2.0211982.2522562.302042
0.50.50.5766320.6706040.692882

The table shows that the converted budget β\beta depends on the prior reconstruction success rate α\alpha. For example, ε=3\varepsilon=3 at α=0.1\alpha=0.1 gives β=2.021\beta=2.021, nearly the same as ε=10\varepsilon=10 at α=10−5\alpha=10^{-5}, which gives β=2.009\beta=2.009. Even at the same prior α=0.1\alpha=0.1, the budgets for ε=3,5,10\varepsilon=3,5,10 are relatively close: 2.021, 2.252, and 2.302. Choosing a prior arbitrarily can therefore obscure differences between DP guarantees. We instead use the prior αε⋆\alpha^\star_\varepsilon at which DP permits its largest reconstruction advantage. With δ\delta fixed, this gives a unique, increasing conversion from ε\varepsilon to β\beta. The resulting comparison matches the posterior-success bounds at that selected prior.

AuthorsOpenReview ↗

Author Response (4/4)

Q6: Run an actual reconstruction attack

First, our theoretical reconstruction bound is attack-independent. Under the stated distribution, reconstruction criterion, and privacy assumptions, it upper-bounds the success probability of any attack observing the release.

We added an extraction attack on FashionMNIST and all 50 CelebA-HQ identities using the approach proposed by Carlini, Nicolas, et al. The attacker accessed only each released model and its training prompts. For each class or identity, the attacker repeatedly generated images and selected the most likely candidates. We then evaluated the attack by measuring their distances to the training images to check the chance of reporduction of training images.

FashionMNIST. We fixed the matching threshold at grayscale pixel L2 <0.1<0.1. The table reports the fraction of 60,000 training images being matched by the generated iamges.

ReleaseTraining images matched
Frozen3.26%
Clean6.11%
DP-SGD4.23%
PAC4.17%

CelebA-HQ. We use Self-Supervised Copy Detection (SSCD) similarity, a measure designed to flag copies or near-duplicates of an image (Pizzi et al., 2022). A score of at least 0.5 flags a pair for visual inspection. The table reports training photos flagged by the attacker’s preselected 5000 candidates across 50 identities.

ReleaseTraining photos flagged
Frozen0%
Clean0.5%
DP-SGD0%
PAC0%

The tested attack falls well below the attack-independent theoretical upper bound. On FashionMNIST, DP-SGD and PAC have similar training-image match rates (4.23% and 4.17%), both lower than clean adaptation (6.11%). On CelebA-HQ, we found no confirmed reconstructions from either private method.

Thank you again for your careful review. We hope the clarified privacy assumptions and additional experiments address your concerns.

Reviewer_axg7OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a PAC-private adaptation framework for latent diffusion models to achieve formal reconstruction privacy without the severe utility degradation associated with differential privacy (DP-SGD). Rather than clipping gradients and injecting noise at every optimization step, the authors train low-dimensional adapters—specifically LoRA or Textual Inversion—and apply a single post-training perturbation using anisotropic Gaussian noise calibrated to the empirical covariance of repeated training runs. For LoRA, they introduce a singular-value canonicalization step that removes scale and sign ambiguities prior to covariance estimation, stabilizing the noise calibration. Across few-shot personalization and general dataset synthesis benchmarks, this approach preserves substantially greater visual fidelity, identity alignment, and downstream classification accuracy than DP-SGD under matched reconstruction privacy budgets

Critical strengths

Paper was easy to read and formulations are straight forward. The ideal of apply perturbations once before weight merging is interesting and worth exploring.

Critical weaknesses
  1. Beyond the mathematical fluffery, provable privacy's core mechanism is to hide leaves amounst trees, to amplify privacy protection by mixing and obfuscating an individual's contribution to a big dataset. Few-shot differential privacy thus sounds like an oxymoron. Experimentally one would obtain better results simply by generating synthetic data from a pretrained model and then applying straight DP-SDG on a mixture of synthetic and real data.
  2. All experiments are missing quantitative comparisons against the un-adaptive pretrained model (perfect privacy by construction) and the non-private oracle (ideal fit by construction). Qualitative results show severe artifacts in generated examples that I have not previously observed with either of the two pretrained models. It is hard to scope the effectiveness of the proposed method without this context
  3. Privacy gurantees are posed per sample while the few-shot datasets contain repeat images of the same identity by construction. Real privacy guarantees would be N times worse if we naively extrapolated privacy protection for an identity that was repeated N times.
Questions for authors

See above

1 comment
AuthorsOpenReview ↗

Author Response

Thank you for recognizing the potential of perturbing the learned adapter once before weight merging. Your review also identified where the experimental evidence and privacy target needed to be made more explicit. In response, we have added frozen pretrained-model and non-private adaptation baselines, evaluated your suggestion of combining public synthetic images with private few-shot training, and clarified that our privacy target is reconstruction of individual reference images, not protection of an identity as a whole. We address each point below.

Concern 1: DP-SGD with public synthetic data in the few-shot setting

We appreciate this practical suggestion. For each of six DreamBooth subjects, we generated 100 class-conditioned images using frozen SD V-1.5 and used them as public auxiliary data. At each update, we combined the privatized gradient from real images with a clipped, unnoised gradient from the synthetic images. We evaluated all methods at ε=10\varepsilon=10 over three training seeds per subject.

MethodDINO-I ↑CLIP-I ↑CLIP-T ↑BRISQUE ↓
Frozen SD1.50.1980.6350.23531.9
Clean0.3760.7040.26827.1
DP-SGD0.1780.6240.24528.1
DP-SGD + synthetic0.1920.6280.24727.8
PAC (Ours)0.3380.6830.25928.4

Adding public synthetic images modestly improves DP-SGD: DINO-I rises from 0.178 to 0.192 and CLIP-I from 0.624 to 0.628. Both subject-fidelity scores nevertheless remain approximately at the frozen-model baseline, whereas PAC reaches 0.338 and 0.683, respectively. DP-SGD with synthetic images achieves a BRISQUE score comparable to PAC (27.8 versus 28.4), but its DINO-I and CLIP-I scores remain below the frozen-model baseline. Thus, adding public synthetic data can modestly improve image quality, but does not resolve the loss of subject fidelity under DP-SGD.

Concern 2: Frozen and clean baselines, and visual artifacts

Thank you for pointing out that these reference points were not sufficiently visible in the original presentation. We have revised the main quantitative tables in the paper for all datasets to include both the frozen pretrained model and the non-private clean adaptation. For a compact comparison, the table below reports their primary metrics with both private methods at ε=10\varepsilon=10. PAC achieves a substantially larger improvement.

Setting (primary metric)FrozenCleanDP-SGDPAC
DreamBooth (DINO-I ↑)0.1470.3880.1490.352
CustomConcept101 (DINO-I ↑)0.1260.4050.1410.353
CelebA-HQ (ISM ↑)0.0010.3410.0020.103
FashionMNIST (TSTR accuracy ↑)26.4%60.5%37.2%43.1%
EuroSAT (TSTR accuracy ↑)23.7%47.6%35.5%39.6%
BloodMNIST (TSTR accuracy ↑)10.4%38.1%14.1%26.4%

Regarding your concern about the qualitative examples, we have revised the figures in the paper to show outputs from the frozen pretrained model and the clean adaptation alongside DP-SGD and PAC. The revised presentation makes it easier to assess the artifacts relative to both reference points.

Concern 3: Real privacy guarantees would be N times worse if we naively extrapolated privacy protection for an identity that was repeated N times.

First, we would like to clarify the distinction between an image-level secret and an identity-level secret. Your concern applies to identity-level protection: if all NN photographs of a person are treated as one protected unit, changing them together is a group-level change and leads to a weaker DP guarantee. In our experiments, the identity is what the adaptation is intended to learn. Our privacy objective is to limit reconstruction of the individual reference photographs used for training, not to prevent the model from generating a recognizable image of the subject. We train and evaluate a separate adaptation for each subject, calibrating within that subject's setting and measuring how well the identity is learned as utility. Accordingly, the NN-record group bound is not the privacy guarantee evaluated here as our experiments concern reconstruction of individual training images.

We thank the reviewer again for the thoughtful feedback. We hope the additional comparisons and clarification of our image-level privacy target address the concerns raised.

Comments on the paper

1 comment
Area_Chair_tMpROpenReview ↗

Dear Reviewers

Since the author-reviewer discussion ends soon, I encourage the reviewer to review the authors' rebuttals and let them know whether their responses are satisfactory and whether any concerns remain.

Thanks for your time and effort in reviewing for ICLR 2027.

Best wishes,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.