acceptodds
Under review as a conference paper at ICLR 2027

Do Not Train on My Data: Unlearnable Data via Contrastive Reverse Distillation

Abstract

We propose **CoReDistill** (_**Co**ntrastive **Re**verse **Distill**ation_), a novel and efficient approach for making image data *unlearnable* from unauthorized generative-model training. Unlike existing methods that iteratively *optimize* each image toward an unlearnability objective, we introduce a sampling-based approach via contrastive decoding that directly *generates* unlearnable samples in a one-shot decoding process. We provide theoretical justification that this sampling procedure produces unlearnable samples, and empirically show that it substantially reduces processing time per image while maintaining strong unlearnability compared with optimization-based baselines. We further observe that the resulting protection transfers across different generative-model variants, providing empirical evidence of cross-model unlearnability. These properties make CoReDistill practical and scalable for real-world data protection involving diverse generative models and large-scale datasets.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
data protection, adversarial attack, image generative models, VLM, copyright protection, privacy protection, fairness, societal considerations

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aiwiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper focuses on optimizing image data "non-trainable" for image generative models.

The problem significance is to prevent protected content, such as celebrity faces or IPs, from training and reproducing by the image generative model, meanwhile preserving the image utility to humans.

The paper poses this problem as "data distillation in a reverse direction of training", but reviewer would summarizes it as adversarial examples training in the image generative model context, and will keep using term "adversarial example" for the remaining of the review.

The paper shows the derivation of their adversarial objectives (image cloaking technique), which is a first-order approximation (1-step update), over the auto-regressive image generative model paradigm.

In the experiment, the paper demonstrated that a surrogate auto-regressive model yields adversarial examples can prevent both auto-regressive model and diffusion-based model from fine-tuning over them, and re-generate the target content.

The major claims are 1) novelty of the introduced optimization algorithm; 2) computational efficiency; 3) transfer across different model architectures

Critical strengths
  • The optimization problem (Eqn. 6), and the derivation of the objective mathematically make sense, and it is concise. The derivation shows how a model parameter updates for unlearning, θ−θtarget\theta - \theta_\text{target}, can connect to a data sample optimization problem, where its form minimizes the inner product of two gradients computed over original data (xx) and the cloaked data (x+δx+\delta), mathematically enforcing two updating directions (gradients) geometrically orthogonal to each other.

  • The proposed method demonstrated cross-architecture transfer. Experiment shows that cloaked example yield over auto-regressive image generative models (emu3) can prevent diffusion model (SDXL) fine-tuning over them. This black-box transferability is valuable.

Critical weaknesses

Method

  • The proposed method is hard-coded with auto-regressive text-to-image generative models. The paper discusses and implements L\mathcal{L} of Eqn. 5, only with auto-regressive score function form, which reads this is a method over-engineered for the auto-regressive image gen. Conceptually, diffusion-based, flow-matching-based image generation all model a score function with different forms , can the proposed method works with those generative paradigms? The paper should includes or discusses that.

Experiment

  • Experiment lacks of model diversity. All experiments in this paper are run on two models, emu3 (auto-regressive model) and SDXL (diffusion model). While paper demonstrated that non-trainable examples optimized over emu3 transfer to prevent fine-tuning for SDXL, it is worth including results on other models beyond just emu3 and SDXL.

  • The paper claims "non-trainable", while supported evidence is only from fine-tuning tasks. The whole paper is articulated with claim "the proposed method yield data examples that image generative models cannot trained with and produce relevant examples". But the actual notion of training discussed in the paper is fine-tuning the pretrained image generative models. Training and Fine-tuning are different objectives and non-fine-tunable cannot be used to claim non-trainable. Reviewer understands that conducing controlled experiments by training generative models over the yield examples from scratch is computationally expensive, but still, current experiment results cannot support that the yield non-trainable examples {xi∗=xi+δix^*_i = x_i + \delta_i} with their original captions {yiy_i} can prevent a model trained from scratch from generating the cloaked data.

  • Image utility preservation is not demonstrated or discussed in the paper. The paper claimed proposed method yield examples that are non-trainable to image generative model, while preserving image utility to humans. However, such a preservation is not demonstrated beyond visually imperceptibility to human (e.g., an optimized image visually look similar to its original version).

  • The evaluation is incomplete that paper currently missing evaluation for preserved generation quality on non-protected concepts. More specifically, all reported metrics are focusing on "how bad is the generation on the protected data, after model is fine-tuned on the cloaked data", but after fine-tuning, do models still function well for general data generation? The paper provides no answer to that. (Motivation-wise, the delivered data-cloaking method does not make sense: if I have to prevent model generate X's faces, I have to optimize cloaked samples of X's faces, then ask the model providers to fine-tune model over them --- how do you make sure models are not damaged after fine-tuning over your data)

Questions for authors
  • After fine-tuning over the cloaked examples, does the fine-tuned model still preserve generation quality over the non-protected concepts?

  • Does the proposed formulation (Eqn. 6) work with diffusion models as surrogate models? If not, any particular reason why it won't

9 comments
AuthorsOpenReview ↗

Response to Reviewer aiwi 1/3

Dear Reviewer aiwi,

Thank you for your thoughtful feedback. We have conducted additional experiments and revised the manuscript to address your concerns.

Before responding point by point, we would like to clarify the motivation behind our work.

CoReDistill aims to help creators and data owners protect their content from unauthorized AI training, without requiring negotiation or cooperation with individual AI developers.

Many creators and data owners have publicly demanded an end to unauthorized AI training, calling for opt-out rights, consent, and compensation [1–6]. Yet litigation and negotiation are costly and time-consuming, especially for individual creators who lack the resources to deal with AI companies worldwide, let alone individual developers who fine-tune models on their work without permission.

As a result, creators have limited practical means to prevent unauthorized training, even after explicitly expressing their objections.

This is precisely what motivates CoReDistill: providing a technical solution that allows data owners to protect their content independently of AI developers.

We hope this clarifies the motivation of our work. Our point-by-point responses follow.

1. Applicability Beyond Autoregressive Models

The proposed method is hard-coded for autoregressive text-to-image models. Can it work with diffusion or flow-matching models?

The reverse-distillation objective in Equation 6 is applicable to any differentiable training loss, including diffusion noise-prediction and flow-matching velocity-regression losses: δ∗=arg⁡max⁡∣δ∣p≤ϵ∇θL(θ,x+δ)⊤(−∇θL(θ,x)).\delta^*= \arg\max_{|\delta|p\leq\epsilon}\nabla_\theta L(\theta,x+\delta)^\top\left(-\nabla_\theta L(\theta,x)\right).

However, directly optimizing this objective requires expensive mixed second-order derivatives, making it difficult to scale to real-world datasets.

This motivates our central contribution: realizing reverse distillation through generation rather than iterative optimization.

While directly optimizing the objective for diffusion or flow-matching models incurs this computational cost, contrastive decoding with autoregressive models provides an efficient, one-shot generation process aligned with the reverse-distillation objective.

Therefore, our methodological reasoning is:

Reverse distillation for data protection → Avoid expensive iterative optimization → Construct samples through generation → Contrastive decoding using autoregressive models.

Autoregression is therefore a deliberate choice enabling efficient generation-based cloaking, not an overengineered design.

2. Model Diversity and Cross-Model Generalization

The experiments mainly involve Emu3 and SDXL. Additional evaluations on other generative models would strengthen the evidence for cross-model transferability.

We agree that broader model evaluation is important. In response, we have expanded our experiments on all 50 CelebA-HQ identities to two additional downstream models: FLUX.1-dev (flow matching, DiT) and Stable Diffusion 1.4 (diffusion, U-Net).

(1) FLUX.1-dev

We expanded FLUX.1-dev evaluation from one identity (Sec 3.2) to all 50, comparing every baseline.

Methodsks ISM ↓dslr ISM ↓
Clean0.64750.6342
FSMG0.62220.6019
ASPL0.60860.5868
AdvDM0.61130.5853
PhotoGuard0.62420.6070
CoReDistill0.60180.5807

FLUX.1-dev generates detectable faces under all conditions (FDFR = 0). CoReDistill achieves the lowest ISM under both prompts, outperforming all baselines and demonstrating cross-model protection.

(2) Stable Diffusion 1.4

We additionally evaluated CoReDistill on SD1.4 across all 50 ids.

Methodsks FDFR ↑sks ISM ↓dslr FDFR ↑dslr ISM ↓
Clean0.00810.50960.01500.3933
CoReDistill0.00440.48510.00560.3641

We found that SD1.4, being an older model with less capable image generation than SDXL and FLUX.1-dev, is particularly sensitive to fine-tuning configurations. Despite extensive hyperparameter tuning during the rebuttal period, we failed to find a configuration that consistently produced high-quality generations, even when fine-tuning on clean images. Consequently, the FDFR results are less consistent than those on newer, more capable models.

Nevertheless, CoReDistill reduces ISM under both prompts, indicating weaker reproduction of the protected identities.

Together with the original results, these experiments demonstrate consistent reductions in identity similarity across three downstream models. The protection effects are particularly clear on the more capable models, SDXL and FLUX.1-dev.

We also recognize the value of further cross-model evaluation. We will continue cross-model evaluations during rebuttal and report new results here; otherwise, we commit to expanding them in the camera-ready version if accepted.

AuthorsOpenReview ↗

Response to Reviewer aiwi 2/3

3. Effectiveness Against Training From Scratch

The current experiments evaluate fine-tuning rather than training from scratch. Can the proposed method prevent models trained from scratch from learning protected data?

Thank you for raising this important question. We have conducted additional training-from-scratch experiments using the 41,600-frame Frozen dataset.

As the reviewer noted, full-scale pretraining would be prohibitively expensive, so we conducted a smaller-scale experiment. Specifically, we randomly initialized the weights of Emu3 and independently trained the model on clean and CoReDistill-cloaked Frozen frames for 3 epochs. Throughout training, we measured the loss on two datasets:

  • Clean training set: The original, uncloaked versions of the frames used for training.
  • Clean held-out evaluation set: Additional Frozen frames not included in the training dataset.

We found that Emu3 trained from scratch on CoReDistill-cloaked images exhibits substantially higher loss on both the clean training set and the held-out evaluation set, compared with the model trained on clean images.

The final-checkpoint results are shown below:

Training dataClean training loss ↓Clean evaluation loss ↓
Clean6.47556.4970
CoReDistill-cloaked6.71696.7314

For the complete training loss trajectories, please refer to Section 3.4 (page 9-10, lines 481–509) of the revised manuscript.

These results indicate that CoReDistill makes the original protected content significantly harder to learn not only during fine-tuning, but also during training from scratch.

4. Preservation of Image Utility

Image utility preservation is not quantitatively demonstrated beyond visual similarity.

We have additionally evaluated the visual fidelity of CoReDistill-cloaked images using PSNR, SSIM [7], LPIPS [8], and ArcFace embedding cosine similarity [9].

The evaluation covers all 50 CelebA-HQ identities, with four images per identity (200 images in total).

MetricResult
PSNR ↑27.86 dB
SSIM ↑0.7968
LPIPS ↓0.1255
ArcFace Cosine Similarity ↑0.8932

These results show that CoReDistill-cloaked facial images retain substantial structural similarity (SSIM), relatively small perceptual differences (LPIPS), and high facial identity similarity (ArcFace), while introducing only moderate pixel-level distortion (PSNR).

5. Generation Quality on Non-Protected Concepts

After fine-tuning on cloaked images, does the model still preserve generation quality on non-protected concepts? How can the method be useful if training on protected data also damages the model's general capabilities?

Evaluating generation quality on non-protected concepts is beyond the scope of this work, as CoReDistill is designed to protect data independently of AI developers' cooperation. We do not expect creators to ask AI developers to fine-tune their models on cloaked data.

This motivation is shared by existing image-cloaking research [10, 12–14], they all primarily focus on how bad is the generation quality of the protected concepts, rather than how well the generation quality of unrelated concepts is preserved.

Glaze [10], particularly, has attracted significant attention from creators and reportedly millions of downloads [11]. Such adoption demonstrates the emergent demand for proactive data protection, while the high computational cost of iterative optimization remains a barrier to large-scale deployment.

CoReDistill thus builds on this established research direction by offering a more efficient and scalable solution.

Preserving general model capabilities while respecting creators' rights is a research question primarily from the perspective of AI developers, involving a broader consideration of both parties' interests. In contrast, our work focuses exclusively on the perspective of data owners, and therefore primarily evaluates how effectively CoReDistill prevents unauthorized learning and reproduction of their protected content.

We hope this explanation clarifies why our evaluation prioritizes protection of the target content rather than generation quality on non-protected concepts.

AuthorsOpenReview ↗

Response to Reviewer aiwi 3/3


We sincerely thank the reviewer again for the constructive feedback, which has helped us strengthen both our experimental evaluation and the presentation of our work.

The central motivation of CoReDistill is to enable creators and data owners to protect their content from unauthorized AI training without requiring negotiation or cooperation with AI developers.

To this end, our work addresses two key challenges: cross-model transferability, protecting against unknown downstream models, and computational efficiency, shifting data cloaking from optimization to generation, making large-scale protection more practical.

We hope our additional cross-model and training-from-scratch experiments, together with the clarification of our motivation, help address the reviewer's concerns.

We believe CoReDistill can bring new insights to the research community while helping creators better protect their work. In light of these additional results and clarifications, we would be grateful if the reviewer would reconsider their assessment.

We would be happy to address any further questions or comments.

Sincerely,

The Authors


[1] “No Artists, No Art”: The Importance of Artists Rights to the genAI Ecosystem. Graphic Artists Guild. February, 2026. https://graphicartistsguild.org/no-artists-no-art/

[2] Statement on the Use of AI. Joan Mitchell Foundation. October, 2026. https://www.joanmitchellfoundation.org/journal/statement-on-the-use-of-ai

[3] Our Joint Statement Calling for Transparency, Fairness and Respect for Creators’ Rights in the Age of AI. Design and Artists Copyright Society (DACS) and Partner Organizations. October 16, 2025. https://www.dacs.org.uk/news-events/joint-statement-on-creators-rights-in-age-of-ai

[4] Protect Artists' Rights: Stop Unlicensed AI Training. Artist Management Association (AMA), via MoveOn. July 2025. https://sign.moveon.org/petitions/protect-artists-rights-stop-unlicensed-ai-training

[5] Require Meta to Give Artists a Real Opt-Out From AI Scraping. Brianna Herder, Change.org Petition. November, 2026. https://www.change.org/p/require-meta-to-give-artists-a-real-opt-out-from-ai-scraping

[6] CODA Issues Written Request to OpenAI Regarding Sora 2. Content Overseas Distribution Association (CODA), Japan. October, 2025. https://coda-cj.jp/en/news/817/

[7] Image quality assessment: from error visibility to structural similarity. Wang et al. IEEE Transactions on Image Processing, 2004

[8] The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. Zhang et al. CVPR, 2018.

[9] ArcFace: Additive Angular Margin Loss for Deep Face Recognition. Deng et al. CVPR, 2019.

[10] Shan et al. Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models. USENIX Security 2023

[11] Tech to protect images against AI scrapers can be beaten, researchers show. The Register. July, 2025. https://www.theregister.com/software/2025/07/11/tech-to-protect-images-against-ai-scrapers-can-be-beaten/1136183

[12] Ven Le et al. Anti-DreamBooth: Protecting users from personalized text-to-image synthesis. ICCV 2023.

[13] Liang et al. Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples. ICML 2023.

[14] Salman et al. Raising the cost of malicious AI-powered image editing. ICML 2023.

Reviewer_6ha3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoReDistill, a data cloaking method designed to protect images from unauthorized training by generative models. Unlike existing approaches that rely on iterative optimization to generate protective perturbations, CoReDistill uses contrastive decoding to construct unlearnable images through a one-shot decoding process. Specifically, it contrasts the token probabilities of a pretrained model and a target-adapted model to induce learning effects opposite to those of the original data, while applying visual-token constraints to preserve image appearance. The authors provide a theoretical justification for the proposed approach and evaluate its effectiveness on facial images and copyrighted creative content. Experimental results demonstrate competitive unlearnability, improved cross-model transferability, and substantially reduced per-image cloaking time compared with existing methods.

Critical strengths
  1. Interesting theoretical formulation. The paper provides an interesting theoretical perspective by connecting reverse distillation with contrastive decoding.
  2. Practical efficiency and cross-model transferability. The proposed method substantially reduces per-image cloaking time compared with optimization-based baselines while demonstrating promising cross-model protection across different generative architectures.
Critical weaknesses
  1. Limited evaluation of cross-model generalization and computational efficiency. Although cross-model transferability and efficiency are central claims, the evaluation primarily relies on Emu3 for constructing cloaked images and a small number of downstream generative models. This limits the evidence for generalization across different architectures and training configurations.
  2. Gap between theoretical and practical. The theoretical derivation relies on a first-order approximation and assumes that the target-adapted model is obtained through a single SGD update. However, the actual implementation involves multiple fine-tuning steps, followed by codebook-distance constraints and random token substitutions.
Questions for authors
  1. Could the authors clarify whether the theoretical connection between contrastive decoding and gradient opposition still holds under the practical training setting?

  2. Could the authors provide empirical evidence, such as gradient cosine similarity between clean and cloaked samples, to verify that the final cloaked images retain the intended gradient-opposition property?

  3. The paper mainly constructs cloaked images using Emu3 and evaluates their transferability to SDXL and FLUX.1-dev. How robust is the proposed protection across different surrogate models, downstream model architectures, and fine-tuning configurations?

  4. Could the authors report the computational cost of this preparation stage?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6ha3 1/2

Dear Reviewer 6ha3,

Thank you for your encouraging assessment and thoughtful feedback. Your comments have helped us identify important areas for improvement, and we have conducted additional experiments and revised the manuscript accordingly.

Please find our point-by-point responses below.

1. Cross-Model Generalization

The evaluation primarily relies on Emu3 for constructing cloaked images and a small number of downstream models. How robust is the protection across different surrogate models, downstream architectures, and fine-tuning configurations?

We appreciate the reviewer's concern regarding the limited cross-model evaluation.

Our goal is to develop a one-to-multiple protection method, where images cloaked using a single surrogate model can protect against multiple downstream models. We agree that evaluating different surrogate models is also an important research direction. Given the limited rebuttal period, we prioritized expanding the evaluation to additional downstream models with different architectures and fine-tuning configurations.

Specifically, we extended our experiments on 50 CelebA-HQ identities from SDXL (diffusion, U-Net, full fine-tuning) to FLUX.1-dev (flow matching, DiT, LoRA) and Stable Diffusion 1.4 (diffusion, U-Net, LoRA).

(1) FLUX.1-dev

We expanded the FLUX.1-dev evaluation from a single identity in Section 3.2 to all 50 identities, including comparisons with all baselines.

Methodsks ISM ↓dslr ISM ↓
Clean0.64750.6342
FSMG0.62220.6019
ASPL0.60860.5868
AdvDM0.61130.5853
PhotoGuard0.62420.6070
CoReDistill0.60180.5807

FLUX.1-dev generates detectable faces under all conditions (FDFR = 0). CoReDistill achieves the lowest ISM under both prompts, outperforming all baselines and demonstrating strong cross-model protection.

(2) Stable Diffusion 1.4

We additionally evaluated CoReDistill on SD1.4 across all 50 identities.

Methodsks FDFR ↑sks ISM ↓dslr FDFR ↑dslr ISM ↓
Clean0.00810.50960.01500.3933
CoReDistill0.00440.48510.00560.3641

We found that SD1.4, being an older model with less capable image generation than SDXL and FLUX.1-dev, is particularly sensitive to fine-tuning configurations. Despite extensive hyperparameter tuning during the rebuttal period, we still failed to find a configuration that consistently produced high-quality generations, even when fine-tuning on clean images. Consequently, the results on SD1.4 are less consistent than those on newer, more capable models such as SDXL and FLUX.1-dev.

Nevertheless, CoReDistill reduces ISM under both prompts, indicating weaker reproduction of the protected identities.

Together with the original SDXL results, these experiments demonstrate consistent reductions in identity similarity across three downstream models—SDXL, FLUX.1-dev, and SD1.4. Notably, CoReDistill shows stronger protection on the more capable models, SDXL and FLUX.1-dev.

We also recognize the value of further cross-model evaluation, as the reviewer suggested. We will continue conducting additional experiments during the remaining rebuttal period and report any new results here. If time does not permit, we commit to expanding these evaluations in the camera-ready version if accepted.

2. Theoretical Connection Under Multi-Step Fine-Tuning

The derivation assumes a single SGD update, whereas the implementation involves multiple fine-tuning steps. Does the theoretical connection still hold?

For simplicity, we assumed a single SGD step on the target data in the derivation. The same theoretical derivation also applies to multiple target-data updates.

With multiple steps, (θtarget−θ)(\theta_{\mathrm{target}}-\theta) is the overall parameter change from fine-tuning on the target data. Assuming that this parameter change remains small in our fine-tuning setting, the same first-order approximation in Eq. (9) gives $s(z)\approx -(\theta_{\mathrm{target}}-\theta)^\top \nabla_\theta L(\theta,z)$ ,

where ∇θL(θ,z)\nabla_\theta L(\theta,z) is the gradient on the cloaked sample (zz).

A higher contrastive score indicates that the update on the cloaked sample moves the model less in the direction of target fine-tuning. This makes it harder for the model parameters to approach θtarget\theta_{\mathrm{target}}, which can help prevent the model from learning the target images (as we show empirically in our response below).

AuthorsOpenReview ↗

Response to Reviewer 6ha3 2/2

3. Empirical Verification of Gradient Alignment

Could the authors provide empirical evidence that the final cloaked images retain the intended gradient-opposition property after applying codebook-distance constraints and random token substitutions?

We apologize that our original wording in the manuscript may have misleadingly suggested that CoReDistill guarantees opposite gradients. More precisely, CoReDistill seeks to generate samples that induce opposing learning effects. We have revised the manuscript accordingly to reflect this distinction.

Interestingly, our additional experiments reveal that the gradients of CoReDistill-cloaked samples are nearly orthogonal to the original learning direction, and that the practical cloaking constraints preserve this property. To investigate this, we generated an additional set of unconstrained cloaked images, removing the codebook-distance constraint and substituting 100% of visual tokens.

Recall that our CelebA-HQ experiments use four reference images to construct θtarget\theta_{\mathrm{target}}, while a disjoint set of four images is used for cloaking. We then measured the cosine similarity between the gradients of constrained and unconstrained CoReDistill-cloaked samples, ∇θlog⁡pθ(z)\nabla_\theta\log p_\theta(z), and two target-learning directions:

  • Surrogate alignment: Cosine similarity with θ−θtarget\theta-\theta_{\mathrm{target}}, where θtarget\theta_{\mathrm{target}} is trained on four reference images.
  • Downstream alignment: Cosine similarity with θ−θdownstream\theta-\theta_{\mathrm{downstream}}, where θdownstream\theta_{\mathrm{downstream}} is independently trained on the four clean images used as cloaking targets.

We report the results averaged across 50 identities here:

MeasurementConstrainedUnconstrained
Surrogate alignment−0.017757−0.002079
Downstream alignment−0.031313−0.001843

All cosine similarities are close to zero, indicating that the cloaked-data gradients are nearly orthogonal to the target-learning directions, both with and without the practical constraints.

These results support the proposed mechanism: training on CoReDistill-cloaked data does not help the model learn the protection target, as their gradients have minimal alignment with the original learning direction. Importantly, the codebook-distance constraints and random token substitutions preserve this misalignment while helping maintain visual similarity.

4. Computational Cost of Target-Model Preparation

Could the authors report the computational cost of the preparation stage?

We have additionally calculated the end-to-end computational cost for our large-scale Frozen experiment involving 41,600 frames.

Training the target-adapted Emu3 model for three epochs takes approximately 9 hours on eight GPUs. Cloaking requires approximately 3 seconds per frame, corresponding to 4.33 hours when distributed across eight GPUs.

Therefore, the total end-to-end cost is approximately 13.33 hours, including target-model preparation. For comparison, extrapolating the measured per-image runtimes of FSMG and ASPL to 41,600 frames, assuming eight-way parallelization, gives:

MethodTime per imageTotal cloaking time
CoReDistill3 s4.33 h
FSMG (100–500 PGD steps)15–72 s21.67 h–4.33 days
ASPL (100–500 PGD steps)110–191 s6.62–11.50 days

These estimates cover cloaking alone; FSMG also requires additional target-model preparation. Thus, CoReDistill requires only 13.33 hours end to end, whereas optimization-based methods can require several days for cloaking alone.

More importantly, as dataset size increases, the substantial savings in per-image cloaking time increasingly outweigh the one-time cost of target-model preparation. This highlights CoReDistill's practical advantage for protecting large-scale datasets.


We sincerely thank the reviewer again for the constructive feedback. Your questions have motivated us to strengthen both our theoretical analysis and experimental evaluation.

We believe the additional results further demonstrate the two central contributions of our work: cross-model transferability and computational efficiency. More broadly, our work introduces a new paradigm for data protection, shifting from optimization to generation.

We hope CoReDistill will bring new insights to the research community while helping overcome major barriers to the practical adoption of data protection technologies, ultimately enabling creators and data owners to better protect their content from unauthorized AI training.

We will continue improving the work and would be happy to address any further questions or suggestions.

Sincerely,

The Authors

AuthorsOpenReview ↗

Additional Cross-Model Results as Promised and a Gentle Follow-Up

Dear Reviewer 6ha3,

As promised in our previous response, we have continued our cross-model evaluations and would like to share one more result. We evaluated Emu3-cloaked images using Qwen-Image-2.1 as an additional downstream model and again observed a clear protection effect.

Methodsks FDFR ↑sks ISM ↓dslr FDFR ↑dslr ISM ↓
Clean0.00000.60620.00000.5920
CoReDistill0.00000.54290.00000.5344

CoReDistill reduces identity similarity under both prompts. Together with our previous results on SDXL, FLUX.1-dev, and SD1.4, this further strengthens the evidence for cross-model transferability, one of the two central contributions of our work alongside computational efficiency.

We sincerely appreciate your encouraging assessment and constructive feedback, which have helped us strengthen our work. In light of these additional results, we would be grateful if you would consider reassessing our work.

As the rebuttal period approaches its end, please let us know if you have any remaining questions or concerns. We would be more than happy to address them.

Thank you again for your time and valuable feedback.

Sincerely,

The Authors

Thanks for your response. I will keep my positive rate. Good luck.

Reviewer_7AP2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • The paper proposes a contrastive reverse distillation mechanism to protect image data against unauthorized generative-model training.
  • Dubbed as COREDISTILL, this approach generates samples through contrastive decoding and reverse distillation.
  • While existing cloaking techniques rely on slow, iterative gradient-based optimization for each image, the proposed method constructs perturbed images whose training gradients oppose those induced by the original clean data.
  • It results in models trained on such a dataset experiencing an unlearning effect with respect to target-specific visual characteristics.
  • The contrastive decoding step bypasses computationally heavy second-order derivative calculations required by gradient optimization methods.
  • The method is evaluated on the CelebA-HQ dataset and 41600 frames extracted from Frozen and Frozen 2, using Emu3, SDXL, and FLUX.1-dev.
Critical strengths
  1. The problem statement is clearly defined. The relationship between contrastive decoding and reverse gradient alignment is intuitively appealing.
  2. Exploration of cross-architecture transfer by generating images through Emu3 and evaluating against SDXL and FLUX.1-dev is insightful.
  3. The reported timing of three seconds per image processing is quite an improvement over existing methods, subject to the condition that target-adapted models already exist.
Critical weaknesses
  1. While the method is referred to as “COREDISTILL” in various places, including the title, the abstract, etc., Table 1,3, and 4, and Figure 6 refer to the method as “RevDistill”. A thorough proofreading before submission is necessary to avoid such mistakes.
  2. Equation 5 refers to a single SGD update, and later equations (Equation 7-10) formulate the objective as maximization of reverse-distillation through first-order Taylor expansion. However, in experiments, the target-adapted model is trained for 100 steps (facial data) and three epochs (frames from Frozen movies). In general, the resulting parameter difference won't be equivalent to a single gradient evaluated at the original parameters.
  3. The paper does not provide any evidence regarding the generated samples actually having oppositely aligned gradients.
  4. The computational cost (Figure 2, right) does not take into consideration the computing and time required for training the target-adapted model. Given the significant computational power used in the experiments (eight NVIDIA H200 GPUs), reporting the computational cost of obtaining the target-adapted model would provide a clearer picture of the efficiency of the proposed method.
  5. Numerical improvements over clean training for SDXL (cross-model unlearn ability results reported in Table 1) are relatively small. The paper does not report confidence intervals or standard deviations across multiple downstream fine-tuning seeds, which would have provided a better understanding of the proposed method's effect.
  6. Results reported in Appendix Table 5 for SDXL show an extremely small relative difference. The paper does not provide any statistical analysis establishing its significance.

Minor

  1. No space between “Figure” and “1” in line 082.
  2. Use of the term “encodeds” in Figure 3.
  3. Both “fine-tuning” and “finetuning” are used throughout the text.
  4. While the text at line 146 talks about minimizing the dot product, Equation (6) uses “arg max”.

Summary/Justification

While the paper addresses an interesting problem statement, the core claim of efficient computational cost is not validated (the 3-second-per-image claim omits the cost incurred during training of the target-adapted model). The reported gains for SDXL are quite small. The paper reports no statistically significant studies and provides no means or standard deviations for the results to support its conclusions. There are minor formatting issues as well.

Questions for authors
  • Please refer to the weaknesses section of the review.
5 comments
AuthorsOpenReview ↗

Response to 7AP2 1/2

Dear Reviewer 7AP2,

Thank you for your constructive feedback, according to which we have conducted additional experiments and revised the manuscript accordingly. We believe this has have strengthened our work. Please find below our point-by-point responses.

1. Terminology and minor corrections

Inconsistent naming of CoReDistill and minor presentation issues.

We have standardized the method name to CoReDistill throughout the paper and corrected the other minor presentation issues.

2. Theoretical justification for multi-step fine-tuning

The theoretical derivation assumes a single SGD update, whereas the experiments involve multiple fine-tuning steps.

For simplicity, we assumed a single SGD step on the target data in the derivation. The same theoretical derivation also applies to multiple target-data updates.

With multiple steps, (θtarget−θ)(\theta_{\mathrm{target}}-\theta) is the overall parameter change from fine-tuning on the target data. Assuming that this parameter change remains small in our fine-tuning setting, the same first-order approximation in Eq. (9) gives $s(z)\approx -(\theta_{\mathrm{target}}-\theta)^\top \nabla_\theta L(\theta,z)$ ,

where ∇θL(θ,z)\nabla_\theta L(\theta,z) is the gradient on the cloaked sample (zz).

A higher contrastive score indicates that the update on the cloaked sample moves the model less in the direction of target fine-tuning. This makes it harder for the model parameters to approach θtarget\theta_{\mathrm{target}}, which can help prevent the model from learning the target images (as we show empirically in our response below).

3. Empirical verification of gradient alignment

The paper provides no empirical evidence that the generated samples have oppositely aligned gradients.

We apologize that our original wording in the manuscript may have misleadingly suggested that CoReDistill guarantees opposite gradients. More precisely, CoReDistill seeks to generate samples that induce opposing learning effects. We have revised the manuscript accordingly to reflect this distinction.

We conducted additional experiments measuring the cosine similarity between the gradients of CoReDistill-cloaked samples, ∇θlog⁡pθ(z)\nabla_\theta\log p_\theta(z), and target-learning directions, θ−θtarget.\theta-\theta_{\mathrm{target}}.

Recall that our CelebA-HQ experiments use four reference images to construct θtarget\theta_{\mathrm{target}}, while a disjoint set of four images is used for cloaking. We evaluate two alignment measures:

  • Target-learning alignment: Cosine similarity with θ−θtarget\theta-\theta_{\mathrm{target}}, where θtarget\theta_{\mathrm{target}} is the target-adapted model trained on the four reference images.
  • Downstream alignment: Cosine similarity with θ−θdownstream\theta-\theta_{\mathrm{downstream}}, where θdownstream\theta_{\mathrm{downstream}} is independently trained on the four clean cloaking-target images.

We report the results averaged across 50 identities here:

MeasurementCosine similarity
Target-learning alignment−0.017757
Downstream alignment−0.031313

Both values are close to zero, indicating that CoReDistill-cloaked gradients are nearly orthogonal to the target-learning directions, for both the target-adapted and downstream models. These findings support the proposed mechanism: training on CoReDistill-cloaked data does not help the model learn the protection target.

AuthorsOpenReview ↗

Response to 7AP2 2/2

4. End-to-end computational efficiency

The reported cloaking time excludes the cost of constructing the target-adapted model.

We report the end-to-end cost for our 41,600-frame Frozen experiment.

Training the target-adapted Emu3 model for three epochs takes approximately 9 hours on eight H200 GPUs. Cloaking requires approximately 3 seconds per frame, corresponding to 4.33 hours when distributed across eight GPUs.

The total end-to-end cost is therefore about 13.33 hours.

For comparison, extrapolating measured FSMG and ASPL per-image runtimes to 41,600 frames with eight-way parallelization gives:

MethodTime per imageTotal cloaking time
CoReDistill3 s4.33 h
FSMG (100–500 PGD steps)15–72 s21.67 h–4.33 days
ASPL (100–500 PGD steps)110–191 s6.62 days–11.50 days

CoReDistill requires approximately 13.33 hours end to end, including target-model adaptation. In contrast, optimization-based methods require several days for cloaking alone, with FSMG incurring additional target-model preparation costs.

In Section 3.2, we previously did not report target-adapted-model preparation time because CoReDistill and FSMG prepare target-adapted models of different sizes—Emu3 (8B) and SDXL (3.5B), respectively—making direct comparisons of preparation time less meaningful.

The large-scale Frozen experiment provides a clearer picture of end-to-end efficiency: as dataset size increases, the substantial savings in per-image cloaking time increasingly outweigh the one-time cost of target-model preparation, demonstrating CoReDistill's scalability and practical efficiency.

5. Cross-model results and statistical significance

The SDXL improvements in Table 1 are relatively small, and confidence intervals or standard deviations across multiple fine-tuning seeds are missing.

The SDXL improvements in Appendix Table 5 are small and lack statistical analysis.

We understand the reviewer's concern about the reliability of the observed cross-model protection. However, conducting multi-seed fine-tuning across all baselines and downstream models would require substantial computational resources. We have therefore expanded our evaluation to two additional downstream models to provide further evidence of CoReDistill's consistent cross-model transferability.

(1) FLUX.1-dev

We extended the FLUX.1-dev evaluation from a single identity in Section 3.2 to all 50 CelebA-HQ identities, comparing CoReDistill against all baselines.

Methodsks ISM ↓dslr ISM ↓
Clean0.64750.6342
FSMG0.62220.6019
ASPL0.60860.5868
AdvDM0.61130.5853
PhotoGuard0.62420.6070
CoReDistill0.60180.5807

FLUX.1-dev generates detectable faces under all conditions (FDFR = 0). Nevertheless, CoReDistill achieves the lowest ISM for both prompts, outperforming all baselines and further demonstrating cross-model protection.

(2) Stable Diffusion 1.4

We also evaluated CoReDistill on SD1.4 across all 50 identities.

SD1.4

Methodsks FDFR ↑sks ISM ↓dslr FDFR ↑dslr ISM ↓
Clean0.00810.50960.01500.3933
CoReDistill0.00440.48510.00560.3641

We found that identity learning with SD1.4 is sensitive to fine-tuning configurations. Despite extensive tuning within the rebuttal period, even clean-data fine-tuning did not consistently produce high-quality generations, which results in the less consistent FDFR results.

Nevertheless, CoReDistill reduces ISM under both prompts, indicating weaker reproduction of the protected identities.

Together with the original SDXL results, these experiments demonstrate consistent reductions in identity similarity across three downstream models—SDXL, FLUX.1-dev, and SD1.4—strengthening the evidence for CoReDistill's cross-model transferability.

Regarding statistical significance, we will continue working on multi-seed fine-tuning and report confidence intervals and standard deviations if computational resources permit during the rebuttal period. Otherwise, we commit to including these analyses in the camera-ready version if accepted.


We sincerely thank the reviewer again for the constructive feedback, which has helped us strengthen our work.

The two central contributions of our work are cross-model transferability and computational efficiency. Our additional cross-model experiments and end-to-end cost analysis strengthen the evidence for both contributions, which we hope will help address the reviewer's concerns about robustness and preparation costs.

We hope this approach will bring new insights to the research community while helping creators and data owners better protect their content from unauthorized AI training. We would be happy to address any further questions or comments.

Sincerely, The Authors

AuthorsOpenReview ↗

Additional Cross-Model Results and a Gentle Follow-Up

Dear Reviewer 7AP2,

We would like to share one more experimental result. To further validate the consistency and robustness of CoReDistill's cross-model protection, we evaluated Emu3-cloaked images on an additional downstream model, Qwen-Image-2.1, and again observed a clear protection effect.

The results are:

Methodsks FDFR ↑sks ISM ↓dslr FDFR ↑dslr ISM ↓
Clean0.00000.60620.00000.5920
CoReDistill0.00000.54290.00000.5344

CoReDistill reduces identity similarity under both prompts, providing further evidence of its cross-model transferability.

As the rebuttal period is approaching its end, we would also like to kindly ask whether you have any remaining questions or concerns. We would be more than happy to address them.

Thank you again for your time and valuable feedback.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.