acceptodds
Under review as a conference paper at ICLR 2027

Efficient Extraction Attacks for Unconditional Diffusion Models

Abstract

Diffusion models (DMs) have become one of the most powerful image generation models. However, the data privacy risks associated with them have not been fully revealed. The data leakage problem for unconditional DMs trained on images of human faces is especially concerning. In this paper, we ask: "Is it possible to extract training data efficiently from the DMs pretrained on human face datasets?" To answer the question, we first formulate the task of efficient extraction attacks in the context of human face images. We first show efficiency of extraction attacks depends on both the speed of generation and per-sample hit probability. By testing mainstream sampling acceleration methods including SDE and ODE solvers, we validate that faster samplers are indeed more efficient in extracting training data, but ODE solvers are still limited by the low per-sample hit probability. Although the SDE solvers are computationally more intensive per image, they are five times more likely to extract a training image with each generation, making them much more efficient than the ODE counterparts. This suggests that hit probability is a more critical determinant of efficiency than generation speed alone. To fundamentally boost sampling efficiency by increasing the hitting rate, we propose a Generation-Filtering-Resampling framework that adapts the distribution of initial noise to concentrate on the high-probability regions in the latent space. Experiments on both a pixel-space DM and a latent-space DM trained on CelebA-HQ and FFHQ validate the efficiency improvement of our method, which is at least twice as efficient as the existing methods in most cases.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
extraction attack, diffusion model, privacy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2pfdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

L017: By testing mainstream sampling acceleration methods including SDE and ODE solvers → maybe its better to name them directly here, as all the methods including the non-accelerated techniques are kind of a SDE/ODE sovlers.

The paper’s text and flow looks to me like a memo, the simplest example is authors try to randomly redefine the acronyms, e.g., DM all over the paper.

Figure.1 maybe the trained-DM ? as the pre-trained makes the authors biased that it have been trained on something else that celeb or ffhq?

Contribution.1 Isn’t just similar to use the some pretrained model like dinov2 or something to measure the similarity? which is already quite established. Have the authors tried to test some generators that were trained on multiple domains? e.g., a sreamlined generator that is trained on multiple domains, now try to extract its training data? I think this is key important experiment that authors need to add, otherwise its too narrow the scope, which makes it a workshop paper might be a better fit.

Its good that authors considered both latent and pixel space diffusion models

Is FID a good metric for the reconstruction quality here that authors used for the quantifying the quality? maybe if you are using the a insightface model its better to use that to quantify?

Table.11, Tabl.12 and Table.3 so comparing the first line and the last line for the all the stages of your method and without the authors methods we gain ~15 minute saving on 8 hour scale? how siginificant is this results? it is acknowledgeable that the authors put this all the timing of the different stages.

 Also, although there is a discussion about the applicability of the conditional generator in the appendix but maybe more experiments are required for this, especially by noting my earlier comment about more streamlined generator consideration.

I consider increasing my score if the additional experiments are provided, especially the multi domain one.

Critical strengths

Please see the summary

Critical weaknesses

Please see the summary

Questions for authors

Please see the summary

2 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thank you for your comments.

Paper Writing

L017: By testing mainstream sampling acceleration methods including SDE and ODE solvers → maybe its better to name them directly here, as all the methods including the non-accelerated techniques are kind of a SDE/ODE sovlers.

The paper’s text and flow looks to me like a memo, the simplest example is authors try to randomly redefine the acronyms, e.g., DM all over the paper.

Figure.1 maybe the trained-DM ? as the pre-trained makes the authors biased that it have been trained on something else that celeb or ffhq?

We will reconsider these naming problem and avoid the ambiguity.

Definition of Extraction

Contribution.1 Isn’t just similar to use the some pretrained model like dinov2 or something to measure the similarity? which is already quite established.

Using pretrained features is a natural approach, and the choice of representation determines which visual content a match preserves. For models trained on faces, our objective is to assess the recovery of facial content while allowing changes in background and illumination. This motivates a representation specifically sensitive to the face.

Our choice followed preliminary comparisons of alternative metrics. When we initially used pixel-space MSE, we observed retrieved pairs with similar black backgrounds and skin tones but visibly different faces. With CLIP and DINO features, we also observed pairs whose similarity was strongly influenced by background and illumination. These observations exposed a mismatch between overall image similarity and the facial content relevant to our task. We additionally observed generated faces resembling training faces under different backgrounds or lighting, which motivates accommodating such variations in the evaluation.

Face-recognition representations provide a well-developed way to emphasize facial identity while reducing sensitivity to these nuisance factors. We therefore use InsightFace features to operationalize extraction in this face-focused setting.

In all, we do not claim the paradigm of using some pretrained model to measure the similarity as one of our contributions. Instead, we provide a more precise way to measure the privacy risk in the context of human faces and we focus on the efficiency problem of extraction attacks, which has been rarely studied.

Experiments on Multiple Domains and conditional generators

Have the authors tried to test some generators that were trained on multiple domains? e.g., a sreamlined generator that is trained on multiple domains, now try to extract its training data? I think this is key important experiment that authors need to add, otherwise its too narrow the scope, which makes it a workshop paper might be a better fit.

Also, although there is a discussion about the applicability of the conditional generator in the appendix but maybe more experiments are required for this, especially by noting my earlier comment about more streamlined generator consideration

The existing non-face experiment in Appendix B evaluates LSUN-Churches using SSCD similarity. Table 6 reports 69 matched training examples in 5.70 hours for SDE-DPM-Solver++ and 195 in 4.97 hours for GFR, corresponding to 12.11 and 39.24 matches/hour. This provides evidence in an additional scene domain. We agree that a generator jointly trained on diverse content would test a broader setting.

We also validate the effectiveness of our method in a conditional diffusion models trained on ImageNet, a multi-domain dataset. Specifically, we use the conditional LDM pretrained on ImageNet as the victim model and evalute the effciency of extraction attack on 10 classes, including insects, birds, flowers and mammals. Note that the extraction is defined by SSCD similarity here following [R3]. We compare GFR with 60-step DPM-Solver and 60-step SDE-DPM-Solver.

SettingTotal generated imagesDistinct matched training imagesTimeEfficiency
DPM-Solver baseline80k3523h 48min92.15/ h
SDE-DPM-Solver baseline80k3663h 49min95.81 / h
GFR80k3883h 33min109.3 / h

GFR extracts about 13.5 more images than SDE-DPM-Sovler, which is also more efficient than its ODE counterpart. Moreover, we find that classes behave quite differently in extraction attacks. For the great grey owl, all three methods are able to extract around 100 images under the 4000 generation budget, while that number for the koala is 0 for all of them. We believe that intra-class diverity is a important factor for memorization and we leave this problem for future research.

AuthorsOpenReview ↗

Rebuttal (2/2)

Pixel-space Diffusion Models

Its good that authors considered both latent and pixel space diffusion models

Thank you for recognizing this aspect of the evaluation. Table 2 includes both latent-space diffusion models and a pixel-space DDPM. We will retain these comparisons and state their data and model scope explicitly.

Image Quality Metric

Is FID a good metric for the reconstruction quality here that authors used for the quantifying the quality? maybe if you are using the a insightface model its better to use that to quantify?

The FID here is not used to evaluate the reconstruction fidelity, but to evaluate the distributional discrepancy between generated and reference image sets. For the reconstruction fidelity, InsightFace similarity already measures the correspondence between each generated face and its matched training image.

Efficiency Analysis

Table.11, Tabl.12 and Table.3 so comparing the first line and the last line for the all the stages of your method and without the authors methods we gain ~15 minute saving on 8 hour scale? how siginificant is this results? it is acknowledgeable that the authors put this all the timing of the different stages.

The main reported benefit is the number of distinct matched training entries obtained within approximately the same runtime. We take Table. 3 for example here.

MethodDistinct matched training entriesReported total timeMatches/hour
60-step SDE-DPM-Solver1288 h 28 min15.12
GFR, combined2368 h 12 min28.78

These reported figures correspond to 16 minutes less runtime, 84.4% more distinct matched entries, and a 1.90× throughput ratio.

[R3] Detecting, Explaining, and Mitigating Memorization in Diffusion Models. ICLR 2024.

Reviewer_SiiBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper provides a method to improve the speed and efficiency for white box attack. the method seems reasonable and the results provided are interesting. more analysis are needed to evaluated the quality of extraction.

Critical strengths
  1. the method proposed is interesting that the generated data are used to train a GMM model for resampling.
  2. it's interesting that SDE achieves better extraction speed and efficiency than ODE.
Critical weaknesses
  1. the author defines the cos sim as the metric to determine the success of the extraction. This definition makes sense but also not convincing. Given the dataset is presented in the experiment. it would be better to provide more directly scores. (such as face recognition, classification, psnr etc...) Otherwise, it's hard to conclude a particular extraction is successful.

  2. in the example of celebrity dataset, the generation is only grouping different photos and is not explicitly covering different photos from the same person. in this case how could the extraction speed be measured? extracting many different images from same person might be easier than extracting different persons' image.

  3. with the method the paper proposes. it's important to understand the diversity from the generation.

Questions for authors
  1. what would be potential limitations of using GFR? are there measured data on this?
2 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Thanks for your comments.

Validation of the extraction criterion

We show that our definition of extraction is also supported by other metric. Our existing criterion uses pretrained InsightFace recognition features. For each generated image, we retain its most similar training image and count a successful match only if their cosine similarity exceeds 0.65. We then deduplicate the matched training images. To provide a complementary measure of visual similarity, we evaluated SSCD on all 162 successful generated–training pairs identified among approximately 100k images produced by the 1000-step DDPM sampler on CelebA-HQ.

The matched pairs have a median SSCD similarity of 0.412, compared with 0.099 for random training-image pairs and 0.336 for pairs between the matched training images and other training photos of the same identity (Table A1). These comparisons provide complementary evidence of visual similarity for the face-matched pairs.

Table A1. SSCD similarity of face-matched pairs and training-image controls.

GroupComparisonsMeanMedian5th percentile95th percentile
Generated image and its face-matched training image1620.4070.4120.2570.554
Matched training image and another training photo of the same identity2,5570.3350.3360.0730.585
Same-identity training-image pairs across the full dataset120,5660.2390.2210.0540.480
Random training-image pairs199,9900.1070.099−0.0120.251

We also directly compared each generated image with the other training photos of its matched identity. The face-selected image scores higher under SSCD in 73.4% of comparisons with these same-identity competitors. We conclude that our face similarity align mostly with SSCD similairity for face-centric images. Because SSCD similarity also takes the non-face part, which is not our focus here, of an image into consideration, they might diverge in some cases.

Image-level and identity-level extraction efficiency

To evaluate extraction speed, for each generated image, we selects only the training image with the highest face similarity, then checks whether its score exceeds the threshold. Then we collect all the matched training images with scores exceeding the threshold and deduplicate them. The speed of extraction is measured by dividing the number of these deduplicated images by the time. Therefore, the speed of extraction is measured on a per-training-image basis.

To show that our method is also efficient in the sense of identities, we show that GFR is not limited in exploiting faces of existing identities. In a seperate rerun of the experiment in Table 3, we additionally evaluate the number of extracted identities. At approximately the same 200k-generation budget, GFR matches 226 distinct training images covering 52 identities, compared with 97 images covering 44 identities for the SDE-DPM-Solver baseline. Identity coverage therefore increases by 18.2%. The Resampling stage contributes 20 identities absent from the Generation stage. Among these, 18 are also absent from the baseline output in this comparison. These results show additional identity coverage in GFR.

Table A2. Image and identity counts in the supplementary CelebA-HQ LDM experiment.

SettingGenerated imagesSuccessful generationsExtracted training imagesNew extracted images vs. stage 1Extracted identitiesNew identities vs. stage 1
SDE-DPM-Solver baseline200k14297—44—
GFR Generation stage100k5948—32—
GFR Resampling stage100k5,0452091783920
GFR combined200k5,104226—52—

For an end-to-end efficiency comparison, we include GFR's required preprocessing. The resulting GFR attack time is 26,715 s (7.42 h), yielding 7.01 identities/h, compared with 5.56 identities/h for the baseline. This is an approximately 26% increase in identity throughput.

AuthorsOpenReview ↗

Rebuttal (2/2)

Diversity of the generated results

We examine both coverage and repeated matching. Table A2 shows that GFR covers more distinct training images and identities at the evaluated generation budget. The Resampling stage adds 178 training-image matches and 20 identity matches beyond the Generation stage, providing evidence that the process continues to discover additional content.

The successful generations are nevertheless strongly concentrated. Resampling produces 5,045 successful generations matching 209 training images, or approximately 24.1 successful generations per matched image. For the complete GFR run, this ratio is 22.6 (5,104/226), compared with 1.46 (142/97) for the baseline.

Broader coverage and greater concentration therefore coexist in this experiment.

Limitations and measured evidence

One limitation of GFR is that the distribution of its generated images is distorted compared to images generated by ordinary samplers. This problem manifests itself mainly in two aspects. First, successful matches concentrate on a subset of training images, as quantified by the repeated-match fractions above. Second, as shown in the original Appendix J, there is also a generation-quality cost. For the CelebA-HQ LDM, the reported FID increases from 22.98 for the baseline to 46.34 for the Resampling stage. However, as shown in Appendix I, we argue that higher FID doesn't affect the quality of the successful generations.

Reviewer_YRe2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how to improve the efficiency of training-data extraction attacks against unconditional diffusion models, with a particular focus on human face datasets. The authors first introduce a face-similarity-based definition of extraction and identify two factors that affect extraction efficiency: generation speed and per-sample hit probability. By comparing different samplers, they observe that stochastic SDE solvers achieve substantially higher extraction rates than deterministic ODE solvers. Building on this observation, the authors propose a Generation-Filtering-Resampling framework. The method first generates a pool of images, selects potentially memorized samples using clique-based filtering, and then fits a Gaussian Mixture Model to their intermediate latents for further sampling. Experiments on LDMs trained on CelebA-HQ and FFHQ, as well as a pixel-space DDPM, demonstrate improved extraction efficiency over existing baselines. Additional ablation studies investigate the contributions of the filtering and resampling stages.

Critical strengths
  1. This paper addresses an important research challenge of privacy attacks against diffusion models: how efficiently an attacker can extract training data under a limited computational budget. I find the distinction between generation speed and extraction success particularly useful, as faster sampling does not necessarily lead to more successful extractions.

  2. The proposed GFR framework is straightforward and does not require retraining the victim model. The experimental results show consistent improvements across the evaluated diffusion models. For example, on the CelebA-HQ LDM, the extraction rate increases from 15.11 to 28.78 images per hour compared with the SDE-DPM-Solver baseline.

Critical weaknesses
  1. My main concern is whether the proposed face-similarity-based criterion reliably measures training data extraction. This work defines an extraction as successful if the cosine similarity between the generated face embedding and that of a training image exceeds 0.65. While I agree that pixel-level similarity can be too restrictive for human faces, face recognition features are primarily designed to capture identity rather than similarity to a specific image. For example, two different photographs of the same person could have high face similarity even when the generated image does not reproduce a particular training example. This distinction matters because the paper interprets a successful identity match as evidence of training data extraction. I think this work would benefit from a more careful validation of this metric, such as evaluating false matches against held-out images, including different images of the same identities, or manually examining a representative set of reported extractions. Since the proposed metric directly determines the reported extraction rates, its reliability is important for assessing the main claim.

  2. Although the experiments demonstrate improvements in extraction throughput, I think the evaluation could more clearly separate the contributions of the learned sampling distribution and the reduced denoising cost. In GFR, resampling starts at an intermediate timestep rather than from the standard Gaussian at the final timestep. This both changes the sampling distribution and reduces the number of denoising steps. The component ablations in Appendix D are helpful, but a comparison against a simpler restart strategy, such as perturbing the filtered intermediate latents without fitting a GMM, would provide a stronger justification for the proposed distribution modeling. Another concern is the generation quality. As reported in Appendix J, the FID on the CelebA-HQ LDM increases from 22.98 for the baseline to 46.34 in the resampling stage. The authors argue that the successful extractions remain visually plausible, but the qualitative examples alone do not fully establish this. It would be useful to report the extraction efficiency under a common generation-quality constraint or provide additional quantitative analysis of the successfully extracted samples. Such results would help clarify the extent to which the reported improvement reflects more effective extraction rather than a trade-off involving generation quality.

Questions for authors
  1. How was the face similarity threshold of 0.65 selected? Have the authors evaluated the false-positive rate using held-out face images, especially different photographs of identities already present in the training set? I would be interested in knowing how many reported extractions remain valid under a more conservative or independently validated criterion.

  2. I am curious if the authors considered a simpler baseline that directly perturbs or adds noise to the filtered intermediate latents, followed by denoising from the same timestep? Such a comparison, under a matched computational budget, would help establish whether fitting a GMM provides an advantage beyond restarting from promising regions.

  3. In Eq. (2), the quantity pˉ\bar{p} is calculated using the number of distinct extracted training images divided by the number of generations. Since repeated extractions are excluded, this quantity can also depend on the total sampling budget and the number of duplicates. Could the authors clarify the distinction between this metric and the probability that an individual generation matches any training example? I think it would be valuable if the authors could show both the total number of successful matches and the number of unique extractions across different sampling budgets.

4 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

Thanks for comment.

Response to Weakness 1 and Question 1

To clarify, our goal is to recover the facial content of training images. We agree that our criterion possibly admits different images of the same identities. However, this problem can be alleviated by tuning the threshold and, in practice, the face-matched training photographs tend to receive higher SSCD scores against the generated images than other training photographs of the same identities.

1. Face similarity, threshold selection, and top-1 matching

We selected 0.65 empirically as an operating threshold that is not too strict in order to reveal privacy risk. We sampled 100 CelebA-HQ identities and constructed three groups of 100 pairs from the training set. Same-photo controls used cropping, resizing, and JPEG compression.

Pair typeMedian face cosine> 0.65> 0.70> 0.75> 0.80
Original and transformed copy0.9906100/100100/100100/100100/100
Same identity, different photographs0.706169/10054/10033/10014/100
Different identities0.00620/1000/1000/1000/100

These distributions show that face similarity captures variation within an identity and that higher thresholds reject more different-photo pairs. They also confirm that 0.65 can accept different photographs of the same person. However, we argue it is not always better to incease the threshold. As shown in Appendix C, a higher threshold means less extractions. When setting the threshold above 0.8, the number of extractions for 1000-step DDPM drops to 0, which actually conceals the privacy risk of the model. It might be the nature of the diffusion models that they can hardly 100% recover a face in the training set.

To clarify, for each generated image, our evaluation selects only the training image with the highest face similarity, then checks whether its score exceeds the threshold. Thus, multiple training photographs exceeding 0.65 contribute at most one match per generation. Changing the threshold leaves the selected training image unchanged whenever the match remains accepted. We further deduplicate the matched training images. This clarifies the counting rule, while the strength of the correspondence requires the additional evidence below.

2. SSCD validation with same-identity competitors

We further evaluated whether the training photograph selected by face similarity is also preferred by SSCD over other photographs of the same person. For each of the 162 reported matches from 1000-step DDPM sampling, we fixed the generated image xx and compared its SSCD similarity to the face-matched training image yy with its similarity to every other training photograph of that identity. We find that the median SSCD similarity between xx and the face-matched yy is 0.412, which is 0.95 higher than that between xx and same-identity alternatives. Moreover, they generally win 73.4% of comparisons with same-identity alternatives. These results provide cross-metric support for the relative similarity of the selected training photographs. However, we must point out that SSCD similarity also takes the non-face parts into consideration. For images in which the human face is the primary subject, SSCD similarity and the face similarity measure used in our work are largely mutually corroborative, with only a small number of cases where they diverge.

AuthorsOpenReview ↗

Rebuttal (2/3)

Response to Weakness 2 (distribution modeling) and Question 2

Following your suggestion, we compared GMM resampling with perturbed restart on the CelebA-HQ LDM. Both strategies use the same 417 filtered images, shared inverted latents, intermediate start time, and remaining 42-step SDE-DPM-Solver schedule, generating 40,000 images each. Perturbed restart uniformly selects a cached latent from the filtered images and adds isotropic Gaussian noise with standard deviation 0.2 times the forward-process noise standard deviation at the restart time. The GMM uses 400 diagonal-covariance components.

Resampling strategyNew extracted training imagesNew matched identitiesSampling time (min)
Perturbed restart1551372.9
GMM1542074.0

Counts use the face-similarity criterion and are deduplicated relative to the shared Generation stage. With the denoising cost controlled, GMM yields essentially the same observed number of new image matches as perturbed restart (154 vs. 155). However, GMM reaches more new identities in this run (20 vs. 13), suggesting a coverage benefit.

Perturbed restart can itself be expressed as a constrained 417-component Gaussian mixture, with uniform weights, centers fixed at the cached latents, and a shared isotropic covariance set by the perturbation strength. This construction requires no mixture fitting. As for GMM, which typically uses fewer components than the number of data points, needs to be fitted before sampling. This process helps cluster similar points and learn the commonnalities of them. We believe that this mechanism can help the model explore more identities.

AuthorsOpenReview ↗

Rebuttal (3/3)

Response to Weakness 2 (generation quality)

We show that successful generations of GFR are of acceptable quality. We evaluated them using two no-reference quality predictors, DeQA-Score [R1] and DSL-FIQA [R2] (higher is better for both). We also include training images and images generated by 1000-step DDPM for comparison.

DeQA-Score. DeQA-Score assesses overall perceptual image quality using a multimodal language model trained to predict human quality ratings [R1]. We select 199 successful generations of GFR, 162 successful generations of 1000-step DDPM and 100 training images for evaluation.

Images GroupnMeanMedian
Resampling matches1993.1153.146
1000-step DDPM matches1623.5533.576
Random real training images1002.9252.909

Notebly, real training images receive lower scores than 1000-step DDPM, consistent with the known preference of no-reference IQA models for the smooth appearance of generated faces. Though the scores of successful generations of GFR are lower than 1000-step DDPM, they are slightly higher than the training images.

DSL-FIQA. DSL-FIQA assesses the perceptual quality of facial images by modeling image degradations and using facial landmarks to emphasize visually salient facial regions [R2]. We additionally scored generated samples using this face-specific quality predictor.

Image groupnMean DSL-FIQA
Resampling matches2000.5220
1000-step DDPM matches1620.6403
Random images from the full DDPM pool2000.6176
Random real training images2000.4808

Again, the scores of successful generations of GFR is lower than 1000-step DDPM but higher than the training images.

Together, we do observe generation quality degradation through quantitative comparison, but the generation quality of these successful generations is still in an acceptable level (close to training samples).

Response to Question 3

Thank you for pointing out this distinction. Suppose we generate KK images in an experiment. Among them, there are HH images that we find to be almost identical to one of the training images. However, there are cases that multiple generated images match the same one training image. This makes the number of unique extracted training images UU to be smaller than HH. Apparently, the metric p^=UK<=HK\hat{p} = \frac{U}{K} <= \frac{H}{K}, where HK\frac{H}{K} is the probability that an individual generation matches any training example.

We take LDM experiment on CelebA-HQ as an example.

SettingGeneration budget KKSuccessful generations HHUnique extracted images UUNew extracted images vs. stage 1
SDE baseline200k187128—
GFR stage 1100k8678—
GFR stage 3100k2175215158
GFR combined200k2261236—

Note that GFR stage 1 is the same as SDE baseline except for the generation budget. Comparing the first two rows, we notice that as the generation budget KK increases, the probability that an individual generation matches any training example remains (from 8.6e−48.6e^{-4} to 9.35e−49.35e^{-4}), but p^\hat{p} drops (from 7.8e−47.8e^{-4} to 6.4e−56.4e^{-5}). The saturation of p^\hat{p} can limit the efficiency of SDE baseline when KK gets even larger. GFR managed to keep the overall p^\hat{p} from dropping.

[R1] Teaching Large Language Models to Regress Accurate Image Quality Scores using Score Distribution. CVPR 2025.

[R2] DSL-FIQA: Assessing Facial Image Quality via Dual-Set Degradation Learning and Landmark-Guided Transformer. CVPR 2024.

I thank the authors for the rebuttal. Most of my concerns are addressed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Final Remarks

Dear AC and Reviewers,

We are grateful for the reviews' engagement in a fruitful discussion and for their recognition of our work. Here we provide a breif summary for the AC and all reviewser.

The reviewers have acknowledged the strengths of our work, including addressing the challenge of extraction attack (YRe2), straightforward and promising method (YRe2, SiiB), interesting findings (SiiB), method effectiveness (YRe2), coverage of model types (2pfd), and clear writing (YRe2). During the rebuttal and discussion phases, the reviewers raised some points of concern, to which we have provided detailed point-by-point responses.

Specifically, our discussions with the reviewers mainly focused on six key aspects:

  • Definition of Extraction (YRe2, SiiB): We evalute the false-positive rates under different thresholds and validates our definition of extraction with SSCD similarity. We show that face similarity reliably establishes training-data extraction in our context.

  • Metric and additional experiment (SiiB, 2pfd): We evaluate both image-level and identity-level extraction in a rerun of our experiment. We demonstrate that GFR is not limited in exploitating existing identities or images. We also validate the effectiveness of GFR on a class-conditional LDM trained on ImageNet.

  • Limitations (YRe2, SiiB): We quatitatively evalaute the image-quality degradation problem of our method. The degradation problem does exists, but it is in an acceptable level.

Once again, we thank all the reviewers and the AC for the efforts in reviewing our work, which are very valuable in improving our work to benefit the wider community. We are committed to reflecting all the clarifications, discussions and comparisons in our revision based on the reviewers' suggestions.

Best regards,

Authors of Submission 22767

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.