acceptodds
Under review as a conference paper at ICLR 2027

Copying as Defense: Protecting Images from Editing by Eliciting Model Copying

Abstract

Image editors built on diffusion transformers (DiTs), such as Flux Kontext, make it possible to turn a single photograph into a convincing manipulated image, including harmful and non-consensual edits. Proactive defenses aim to prevent such manipulations by adding an imperceptible perturbation before an image is shared. Existing defenses for DiT-based in-context editors take an intuitive approach: make the output attend less to the protected reference image. But this leaves the model free to follow the edit instruction using the text and its own generated tokens, often producing a new, weakly grounded image rather than preventing the edit altogether. We find that the opposite strategy is more effective. Instead of leading the model to ignore the reference, we encourage it to copy it. Through controlled attention interventions, we show that concentrating target-token attention on the reference image in only a few early blocks is sufficient to collapse the editing behavior, causing the model to reproduce its input, often as a lossy copy, rather than execute the requested transformation. Based on this insight, we introduce Copying-as-Defense (CaD), which maximizes per-token target-to-reference attention mass in early transformer blocks. Since the attention rows are normalized, increasing reference attention suppresses the attention allocated to the text and generated canvas. Restricting the objective to early blocks also lets us truncate backpropagation path, keeping optimization cheap. Across editors and benchmarks, CaD substantially suppresses instruction-following. It reduces average edit success from 81.4% on unprotected images to 24.0%, a 70.6% relative reduction. Compared with DeContext, the strongest baseline on average at 43.0% edit success, CaD further reduces edit success by 44.3%. CaD is also substantially cheaper to optimize, running roughly 1.8–2.0 times faster per step than comparable optimization-based methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, Generative AI, Data Poisoning, Data Protection

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

This paper introduces Copying-as-Defense (CaD), a protection method against DiT-based in-context image editors, which maximizes per-token target-to-reference attention mass in early transformer blocks. In experiments on 50 images with two editors, CaD lowers the average edit success rate from 81.4% to 24.0%, and it is cheaper to optimize, running roughly 1.8–2.0× faster per step than comparable optimization-based methods.

Critical strengths
  1. The idea is simple and well motivated. The perturbation is optimized with an empty prompt, so it does not need to know the attacker's instruction.

  2. The proposed method is cheaper and faster to optimize, which gives it good practical potential.

  3. The results on FLUX.1 Kontext are strong, e.g., edit success drops to 5.1% on BodyBench, compared with 24.9% for DeContext.

Critical weaknesses

Major

  1. The paper does not evaluate robustness to common image transformations such as JPEG compression, resizing and noise. The authors mention this in the Limitations, but I think this evaluation is necessary for a protection method.

  2. The evaluation is white-box and same-model only: each perturbation is optimized and tested on the same editor. Also, all experiments use 512×512 inputs, while these editors are often run at higher resolution.

  3. There is no human evaluation. Edit success is judged by a single VLM (Qwen3-VL-30B-A3B). Also, 50 images is a small test set.

  4. Copying is only clearly shown on Flux Kontext. On FIBO, CaD outputs look like degraded images with blocky artifacts rather than copies (Fig. 3).

  5. CaD uses 1000 PGD steps while the baselines use 800, so the comparison is not fully fair (Fig. 5 only covers Flux Kontext on BodyBench). Also, the per-token objective is motivated by the token-coverage argument, but I didn't find an ablation against an image-averaged version.

Minor

  1. Some notation is inconsistent. For example, in the main text C/T/R denote text/target/reference tokens, but in the appendix T/Z/C denote text/target/context tokens.

  2. The paper uses a lots phrasing such as "rather than ..." (10 times in the main text) and "not ... but ...", which hurts readability.

  3. The citation for unauthorized personalization (Marchal et al., 2024, line 43-44) is a paper about AI misuse, which seems to be a wrong citation.

Questions for authors
  1. How robust is CaD to JPEG compression, resizing and simple purification (e.g., noisy upscaling)?
  2. Does a perturbation optimized on Flux Kontext transfer to FIBO Edit, or to other editors?
  3. If the outputs are copies, why is CaD's ArcFace so low (0.27 on Flux BodyBench)? What is the ArcFace similarity between the original and the protected input?
  4. How well does the Qwen3-VL judge agree with humans?
  5. How does the per-token objective compare with simply maximizing the average reference attention?
5 comments
AuthorsOpenReview ↗

W1. Missing Robustness Evaluation & W2. White-box Evaluation, Cross-model Transferability, and Image Resolution

Dear Reviewer HSBm,

Thank you for your thoughtful review and for recognizing the simplicity, motivation, empirical effectiveness, and computational efficiency of CaD. We appreciate your constructive feedback. We have conducted additional experiments and updated the manuscript to address your concerns, as detailed below.

W1. Missing Robustness Evaluation

Thank you for highlighting this concern. To address it, we conducted additional robustness experiments evaluating cropping, Gaussian noise, and JPEG compression, with results presented below.

Consistent with prior work on adversarial image protection, we find that these transformations substantially weaken protection effectiveness. Following purification, all methods exhibit broadly comparable performance across most metrics, although DeContext achieves lower ESR under cropping. Notably, DeContext's CLIP directional similarity and PickScore remain comparable to those of other methods, suggesting that editing may still occur in a decontextualized manner that is not counted as successful by the VLM. These findings highlight a shared limitation of existing protection methods. Improving robustness through Expectation over Transformations (EoT) represents an important direction for future work.

We have updated the paper and included these robustness results and discussion in the appendix.

Table: Purification robustness at ϵ=12/255\epsilon = 12/255 against FLUX Kontext. None denotes no purification. Lower is better for all metrics.

BodyBench

MetricMethodNoneCropGauss.JPEG
ESR ↓DeContext0.240.420.710.71
DiffPGD0.440.570.760.79
PhotoGuard0.810.760.780.82
Ours0.050.740.790.81
CLIPdir ↓DeContext0.270.260.270.28
DiffPGD0.250.270.300.30
PhotoGuard0.330.300.320.32
Ours0.030.290.310.32
PickScore ↓DeContext22.2921.6121.7821.78
DiffPGD21.3821.7322.0522.05
PhotoGuard22.2821.9322.2522.28
Ours18.6621.9322.1722.24

InpaintBench

MetricMethodNoneCropGauss.JPEG
ESR ↓DeContext0.210.410.640.60
DiffPGD0.520.480.730.68
PhotoGuard0.780.760.790.79
Ours0.160.760.780.80
CLIPdir ↓DeContext0.200.200.210.21
DiffPGD0.190.200.220.22
PhotoGuard0.240.220.230.24
Ours0.050.220.240.24
PickScore ↓DeContext21.8420.1920.5120.58
DiffPGD20.4820.4020.7020.72
PhotoGuard20.8920.7220.9320.84
Ours18.6520.6620.9120.88

W2. White-box Evaluation, Cross-model Transferability, and Image Resolution

First, white-box optimization is an established threat model in proactive image protection, shared by PhotoGuard, DiffVax, DiffusionGuard, DeContext, and AdvPaint. Our experiments target publicly available, open-weight editing models, for which white-box optimization is feasible without model-provider cooperation. CaD protects images before publication and does not require modifying the editing model.

To address your concern, we additionally evaluated cross-model transferability, testing whether perturbations optimized on FLUX Kontext remain effective against other editors without reoptimization or access to their parameters.

MethodLongCat-EditDreamOmni2InstructPix2Pix
Unprotected87.658.452.7
PhotoGuard88.256.443.3
DeContext86.023.810.7
DiffPGD86.944.922.7
Ours86.40.935.1

The results demonstrate that CaD can transfer to unseen editing models, although effectiveness depends on the target architecture. Notably, CaD achieves strong transfer to DreamOmni2 (58.4% → 0.9% ESR), which is built upon FLUX Kontext through LoRA fine-tuning. This suggests that CaD's protection can remain effective across fine-tuned variants of the source model, even without access to their adapted parameters. CaD also provides moderate protection on InstructPix2Pix (52.7% → 35.1%), but all methods exhibit limited transfer to LongCat-Edit, highlighting the challenges of cross-architecture generalization.

Third, regarding resolution, we standardize inputs to 512×512 to ensure consistent experimental conditions and computationally feasible comparisons across methods, following the common evaluation setting in prior optimization-based protection work.

We have updated the paper to clarify these deployment assumptions and resolution limitations, and included the cross-model transferability results in the appendix.

AuthorsOpenReview ↗

W3. Limited Evaluation Scale and Lack of Human Validation

W3. Limited Evaluation Scale and Lack of Human Validation

Thank you for raising this concern. We agree that human validation and statistical analysis strengthen the evaluation. However, our evaluation scale is consistent with established proactive image protection benchmarks, particularly given the computational cost of per-image optimization.

First, our scale is comparable to prior work. AdvPaint evaluates 100 images, DiffusionGuard introduces a benchmark with 42 base images, DiffVax evaluates 75 images, and DeContext uses 50 identities. Our evaluation spans 50 distinct images, 15 editing instructions per image, and two editing models, yielding 1,500 edited outputs per method, alongside multiple baselines and ablations. Each image requires separate optimization for every defense and model (approximately 16–20 minutes for several baselines), making larger-scale controlled comparisons computationally expensive.

Second, our benchmark selection specifically targets a gap in existing evaluations. Many prior benchmarks emphasize face-centric portraits, whereas our central finding concerns protection failures on full-body images, where identity-critical regions occupy substantially fewer reference tokens. We therefore combine face-centric images from InpaintGuardBench with full-body images from BodyBench to evaluate both settings.

Third, our evaluation does not rely exclusively on a single VLM judge. Alongside Qwen3-VL, we report PickScore, CLIP directional similarity, ArcFace similarity, and qualitative comparisons, consistent with established evaluation practices.

To further validate the automated judgments and address your concern, we conducted an independent human evaluation involving 76 participants and 500 edited outputs (50 source images × 2 editing models × 5 methods), with each output evaluated by at least three annotators. Participants viewed the original and edited images side by side, first assessing visual similarity, identity preservation, distortion, and usability without seeing the editing instruction, and then rating whether the requested edit was successfully performed on a five-point Likert scale. We report the percentage of ratings indicating agreement or strong agreement that the edit was performed, along with 95% bootstrap confidence intervals obtained by resampling source images.

Table: VLM ESR versus Human Editing Success

EditorMethodPaper VLM ESR (%) ↓Matched VLM ESR (%) ↓Human Edit Made (%) ↓ [95% CI]
FLUX KontextUnprotected81.886.093.3 [88.0, 97.3]
PhotoGuard80.380.086.0 [78.7, 92.7]
Diff-PGD47.350.074.7 [64.7, 84.0]
DeContext23.320.071.3 [62.0, 80.0]
Ours9.512.011.3 [4.0, 19.3]
FIBO EditUnprotected81.182.090.0 [83.3, 96.0]
PhotoGuard81.776.085.3 [78.0, 92.0]
Diff-PGD49.638.056.7 [45.3, 68.0]
DeContext63.764.077.3 [68.0, 86.0]
Ours41.248.050.0 [39.3, 60.7]

Table: Comparison of VLM-based and human-evaluated editing success rates. Paper VLM ESR is computed on the full benchmark, whereas Matched VLM ESR uses the subset selected for human evaluation. Human editing success measures whether the requested edit was performed, with 95% bootstrap confidence intervals. Each editor–method condition contains 50 edited outputs. Unlike the human metric, VLM ESR additionally requires identity preservation and the absence of major visual artifacts. Lower is better.

Importantly, the human evaluation reinforces our central motivation. For DeContext, we observe a substantial discrepancy between VLM ESR and human-judged instruction fulfillment. On FLUX Kontext, only 20.0% of outputs satisfy the VLM's stricter success criteria, yet humans identify the requested edit in 71.3% of outputs. This suggests that DeContext can still permit the requested transformation while disrupting the original image's identity or context, causing the VLM to classify these outputs as unsuccessful edits. This motivates CaD's strategy of inducing copying or output corruption. In contrast, CaD's VLM and human success rates closely agree (12.0% vs. 11.3% on FLUX Kontext; 48.0% vs. 50.0% on FIBO Edit), providing additional support for its effectiveness.

We have updated the paper to clarify our benchmark selection and included the human evaluation and uncertainty estimates in the appendix.

AuthorsOpenReview ↗

W4. Copying Behavior on FIBO Edit and Low ArcFace Similarity & W5. Unequal Optimization Steps and Missing Per-token Ablation

W4. Copying Behavior on FIBO Edit and Low ArcFace Similarity

Thank you for raising this concern. We agree that the blocky artifacts observed on FIBO Edit make the copying interpretation less apparent. Our use of copying refers to the model reproducing the reference image's content and spatial structure instead of executing the requested edit, not necessarily producing a pixel-perfect reconstruction.

To directly validate this mechanism, we conducted an additional evaluation measuring whole-image similarity between generated outputs and the original clean reference, aggregated across both benchmarks on FLUX Kontext.

MethodDINO ↑R@1 (%) ↑LPIPS ↓SSIM ↑MS-SSIM ↑PSNR ↑
Unprotected.55185.3.458.528.46111.8
PhotoGuard.49981.6.489.546.50311.8
Diff-PGD.49576.2.621.329.4239.4
DeContext.22429.2.718.370.2329.4
CaD (ours).74091.2.418.535.70017.1

CaD achieves the strongest overall reference preservation, including substantially higher DINO similarity and 91.2% reference retrieval accuracy. Together with the controlled attention interventions in Fig. 2, these results support reference reconstruction rather than arbitrary generation failure on FLUX Kontext.

Regarding FIBO Edit, we clarify that CaD can induce either lossy reference reconstruction or output corruption, depending on the editor and input. The blocky artifacts in Fig. 3 are therefore not necessarily evidence of faithful copying, and we do not claim otherwise.

This also helps contextualize the low ArcFace similarity. Preserving global image structure does not guarantee preservation of fine-grained facial identity. Blurring and distortions in reconstructed outputs can substantially reduce ArcFace similarity even when scene-level content is preserved. Our additional measurements compare generated outputs against the clean reference; we have not separately established the original-to-protected-input ArcFace similarity in this analysis.

We have updated the paper to clarify our terminology as "lossy reference reconstruction," distinguish copying from output corruption, and include the additional quantitative validation in the appendix.

W5. Unequal Optimization Steps and Missing Per-token Ablation

CaD's performance advantage does not arise simply from using more optimization steps. As shown in Fig. 5, on FLUX Kontext with BodyBench, CaD achieves strong protection within approximately 300 steps, outperforming all the baselines' final 800-step results across the reported metrics.

Furthermore, despite using 1,000 steps compared with 800 for the baselines, CaD requires only 10.78 minutes per image, compared with 16.43 minutes for DeContext and 16.60 minutes for Diff-PGD. This reflects its more efficient optimization, requiring only 0.647 s/step versus 1.232 and 1.245 s/step, respectively.

Regarding the per-token objective, we clarify that our token-coverage argument primarily motivates the limitations of DeContext. Our central contribution is the observation that increasing target-to-reference attention induces lossy copying, as demonstrated by our controlled interventions in Fig. 2 and experiments. Indeed, maximizing image-averaged reference attention could also encourage this behavior. Our per-token formulation additionally penalizes uneven reference-attention allocation across target queries, encouraging more uniform reliance on the reference. However, the copying mechanism does not fundamentally depend on the per-token formulation, and we do not claim that it is necessary for inducing copying. We agree that an ablation against image-averaged attention maximization would help isolate its contribution, and we will include this discussion and comparison in the revised paper.

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

This paper proposes Copying-as-Defense (CaD), an image protection framework that hinders unauthorized DiT-based editing by encouraging the model to copy the input image rather than execute edit instructions. The authors observe that suppressing reference-image attention can weaken grounding in the source content without preventing instruction following. Controlled attention interventions show that increasing reference attention in early transformer blocks can instead induce copying behavior. CaD exploits this observation by optimizing small per-image perturbations that increase target-to-reference attention for each target token, reducing the attention mass available to text and target tokens. During optimization, the objective requires computation only through the selected early blocks, reducing computational cost. CaD also operates without requiring a pool of candidate editing instructions. Evaluated on FLUX Kontext and FIBO Edit, it reduces the average edit success rate from 81.4% to 24.0% while shortening optimization time compared with DeContext and Diff-PGD.

Critical strengths
  1. The paper offers a distinctive and well-motivated insight that increasing reference-image attention can suppress editing by inducing copying behavior. Controlled attention interventions compare reference suppression with amplification, providing a concrete experimental basis for the design. Additional ablations examine the effects of objective direction and layer selection.

  2. Restricting the objective to early transformer blocks allows both forward computation and backpropagation to stop before the deeper layers. This design reduces per-step computational cost and total protection time compared with DeContext and Diff-PGD.

  3. CaD does not require a pool of candidate editing instructions during perturbation optimization, yet demonstrates protection against multiple editing requests. This is practically useful because image owners generally cannot anticipate the instructions that will later be applied to their images.

Critical weaknesses
  1. Although the paper evaluates two editing models, perturbations are optimized separately for each target editor. Cross-model transferability therefore remains unevaluated: the experiments do not establish whether an image protected against one editor remains protected when processed by another. Applying the same protected images to unseen editors without re-optimization, alongside corresponding baseline comparisons, would better establish the defense’s practical scope when the target model is unknown.

  2. The paper does not evaluate robustness after common image transformations, such as JPEG compression, resizing, additive noise, or purification-based perturbation removal. Although the authors acknowledge this limitation, these operations are particularly relevant to images distributed publicly. Evaluating protection after such processing would help determine whether the method remains effective under realistic sharing and attack conditions.

  3. The evaluation of the optimized perturbations does not sufficiently distinguish copying-induced protection from general output corruption. Under the ESR definition, identity loss and major visual artifacts also count as editing failures, and artifact-laden outputs are observed for FIBO. Although ArcFace similarity is reported, the paper does not directly quantify whole-image copying fidelity. Reporting perceptual and structural similarity between the source and final output, together with human evaluation of copying versus corruption, would better substantiate the proposed mechanism.

Questions for authors
  1. How well do perturbations optimized for one editor transfer to unseen editors without re-optimization, compared with baseline methods under matched conditions?

  2. How robust is CaD to JPEG compression, resizing, additive noise, and purification-based perturbation removal?

  3. Can whole-image source-to-output similarity measurements and human evaluation verify whether the optimized perturbations induce copying rather than general output corruption?

2 comments
AuthorsOpenReview ↗

W1. Cross-model Transferability & W2. Robustness to Image Transformations and Purification

Dear Reviewer QpsW,

Thank you for your thoughtful review and for recognizing the novelty, motivation, empirical effectiveness, and computational efficiency of our approach. We also appreciate your recognition of CaD's prompt-agnostic protection. We have conducted additional experiments addressing your concerns about cross-model transferability, robustness, and copying behavior, and updated the manuscript accordingly.

W1. Cross-model Transferability

We agree that cross-model transferability is an important consideration for practical deployment. To address this concern, we additionally evaluated cross-model transferability, testing whether perturbations optimized on FLUX Kontext remain effective against other editing models without reoptimization or access to their parameters.

MethodLongCat-EditDreamOmni2InstructPix2Pix
Unprotected87.658.452.7
PhotoGuard88.256.443.3
DeContext86.023.810.7
DiffPGD86.944.922.7
Ours86.40.935.1

The results show that CaD can transfer to unseen editing models, although effectiveness depends on the target architecture. Notably, CaD achieves particularly strong transfer to DreamOmni2 (58.4% → 0.9% ESR), which is built upon FLUX Kontext through LoRA fine-tuning. This suggests that CaD's protection can remain effective across fine-tuned variants of the source model, even without access to their adapted parameters. CaD also provides moderate protection on InstructPix2Pix (52.7% → 35.1%), but exhibits limited transfer to LongCat-Edit, highlighting the challenges of cross-architecture generalization for all methods. These findings support deployment against adapted versions of accessible models, while improving broader cross-model transferability remains an important direction for future work.

We have updated the paper and included the transferability results and discussion in the appendix.

W2. Robustness to Image Transformations and Purification

Thank you for highlighting this concern. To address it, we conducted additional robustness experiments evaluating cropping, Gaussian noise, and JPEG compression, with results presented below.

Consistent with prior work on adversarial image protection, we find that these transformations substantially weaken protection effectiveness. Following purification, all methods exhibit broadly comparable performance across most metrics, although DeContext achieves lower ESR under cropping on both datasets. Notably, while DeContext achieves a lower ESR under cropping, its CLIP directional similarity and PickScore remain comparable to those of other methods. This suggests that editing may still occur, but in a decontextualized manner that is not counted as a successful edit by the VLM, rather than indicating consistently stronger robustness across metrics.

These findings highlight a shared limitation of existing protection methods. Improving robustness through techniques such as Expectation over Transformations (EoT) represents an important direction for future work. We acknowledge that resizing and more adaptive purification strategies, such as noisy upscaling, require further evaluation.

We have updated the paper and included these additional robustness results and discussion in the appendix.

Table: Purification robustness at ϵ=12/255\epsilon = 12/255 against FLUX Kontext. None denotes no purification. Lower is better for all metrics.

BodyBench

MetricMethodNoneCropGauss.JPEG
ESR ↓DeContext0.240.420.710.71
DiffPGD0.440.570.760.79
PhotoGuard0.810.760.780.82
Ours0.050.740.790.81
CLIPdir ↓DeContext0.270.260.270.28
DiffPGD0.250.270.300.30
PhotoGuard0.330.300.320.32
Ours0.030.290.310.32
PickScore ↓DeContext22.2921.6121.7821.78
DiffPGD21.3821.7322.0522.05
PhotoGuard22.2821.9322.2522.28
Ours18.6621.9322.1722.24

InpaintBench

MetricMethodNoneCropGauss.JPEG
ESR ↓DeContext0.210.410.640.60
DiffPGD0.520.480.730.68
PhotoGuard0.780.760.790.79
Ours0.160.760.780.80
CLIPdir ↓DeContext0.200.200.210.21
DiffPGD0.190.200.220.22
PhotoGuard0.240.220.230.24
Ours0.050.220.240.24
PickScore ↓DeContext21.8420.1920.5120.58
DiffPGD20.4820.4020.7020.72
PhotoGuard20.8920.7220.9320.84
Ours18.6520.6620.9120.88
AuthorsOpenReview ↗

W3. Distinguishing Copying from Output Corruption

W3. Distinguishing Copying from Output Corruption

We agree that our original analysis did not directly quantify reference copying. We conducted an additional evaluation measuring whole-image similarity between edited outputs and the original clean reference (not the perturbed image), aggregated across both benchmarks on FLUX Kontext.

MethodDINO ↑R@1 (%) ↑LPIPS ↓SSIM ↑MS-SSIM ↑PSNR ↑
Unprotected.55185.3.458.528.46111.8
PhotoGuard.49981.6.489.546.50311.8
Diff-PGD.49576.2.621.329.4239.4
DeContext.22429.2.718.370.2329.4
CaD (ours).74091.2.418.535.70017.1

CaD achieves the strongest reference preservation, including substantially higher DINO similarity. Its 91.2% reference retrieval accuracy further confirms that outputs retain reference-specific information. Together with the controlled attention interventions in Fig. 2, these results support reference reconstruction rather than arbitrary generation failure on FLUX Kontext.

Regarding FIBO Edit, we agree that the blocky artifacts make the copying interpretation less visually apparent. We clarify that reference copying need not imply identity-faithful or pixel-perfect reconstruction. Our use of the term copying refers to the model's tendency to reproduce the reference image's content and spatial structure instead of carrying out the requested edit. CaD can induce lossy reference reconstruction or output corruption, depending on the editor and input. We do not claim that every artifact-heavy FIBO output constitutes a copy.

To further validate the automated judgments and address your concern, we conducted an independent human evaluation involving 76 participants and 500 edited outputs (50 source images × 2 editing models × 5 methods), with each output evaluated by at least three annotators. Participants viewed the original and edited images side by side, first assessing visual similarity, identity preservation, distortion, and usability without seeing the editing instruction, and then rating whether the requested edit was successfully performed on a five-point Likert scale. We report the percentage of ratings indicating agreement or strong agreement that the edit was performed, along with 95% bootstrap confidence intervals obtained by resampling source images.

Table: VLM ESR versus Human Editing Success

EditorMethodPaper VLM ESR (%) ↓Matched VLM ESR (%) ↓Human Edit Made (%) ↓ [95% CI]
FLUX KontextUnprotected81.886.093.3 [88.0, 97.3]
PhotoGuard80.380.086.0 [78.7, 92.7]
Diff-PGD47.350.074.7 [64.7, 84.0]
DeContext23.320.071.3 [62.0, 80.0]
Ours9.512.011.3 [4.0, 19.3]
FIBO EditUnprotected81.182.090.0 [83.3, 96.0]
PhotoGuard81.776.085.3 [78.0, 92.0]
Diff-PGD49.638.056.7 [45.3, 68.0]
DeContext63.764.077.3 [68.0, 86.0]
Ours41.248.050.0 [39.3, 60.7]

Table: Comparison of VLM-based and human-evaluated editing success rates. Paper VLM ESR is computed on the full benchmark, whereas Matched VLM ESR uses the subset selected for human evaluation. Human editing success measures whether the requested edit was performed, with 95% bootstrap confidence intervals. Each editor–method condition contains 50 edited outputs. Unlike the human metric, VLM ESR additionally requires identity preservation and the absence of major visual artifacts. Lower is better.

Importantly, the human evaluation reinforces our central motivation. For DeContext, there is a substantial discrepancy between VLM ESR and human-judged instruction fulfillment. On FLUX Kontext, only 20.0% of outputs satisfy the VLM's stricter success criteria, yet humans identify the requested edit in 71.3% of outputs. This suggests that DeContext can still permit the requested transformation while disrupting the original image's identity or context, causing the VLM to classify these outputs as unsuccessful edits.

This is precisely the limitation discussed in our paper: decontextualization does not necessarily prevent editing, motivating CaD's strategy of inducing copying or output corruption instead. In contrast, CaD's VLM and human success rates closely agree (12.0% vs. 11.3% on FLUX Kontext; 48.0% vs. 50.0% on FIBO Edit), providing additional support for its effectiveness in preventing requested transformations.

We have updated the paper to clarify our terminology as "lossy reference reconstruction," distinguish copying from output corruption, and include the additional quantitative copying analysis and human evaluation results in the appendix.

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

This paper proposes Copying-as-Defense (CaD), a proactive image protection method that maximizes target-to-reference attention in early DiT blocks to encourage image copying rather than instruction-following editing. Experiments on FLUX Kontext and FIBO Edit demonstrate improved protection effectiveness and computational efficiency over existing baselines.

Critical strengths
  1. Novel perspective. The idea of encouraging reference copying instead of suppressing reference attention is interesting and counterintuitive.

  2. Strong empirical results. CaD substantially reduces editing success rates compared with existing defenses, particularly on FLUX Kontext.

Critical weaknesses
  1. Insufficient validation of the copying mechanism. The paper lacks direct quantitative evidence that protected outputs actually copy the reference image rather than simply exhibit generation failures. The low ArcFace similarities in some settings further raise concerns about this interpretation.

  2. Unclear justification for the per-token objective. Maximizing each target query's total attention to reference tokens does not necessarily ensure sufficient attention to identity-critical regions. Thus, the claimed solution to the token-coverage problem is not fully justified.

  3. While the paper explains the advantage of maximizing reference attention over suppressing it, it does not sufficiently justify why attention manipulation is preferable to alternative objectives, such as feature-targeted optimization or direct copying losses. Additional comparisons would strengthen the motivation for the proposed objective.

  4. Missing robustness evaluation. The paper does not evaluate common image transformations such as JPEG compression, resizing, or noise, nor adaptive purification attacks, which limits its practical applicability.

  5. CaD requires full access to the target editing model and performs model-specific, per-image optimization, which limits its practical applicability. The paper does not clearly discuss how such assumptions can be satisfied in realistic deployment scenarios.

  6. The paper overlooks relevant model-side protection approaches, such as GuardDoor (Zeng et al., 2025) and PersGuard (Liu et al., 2026). If cooperation with model providers is considered a practical deployment scenario, these approaches should be discussed.

  7. Limited evaluation scale. The experiments involve only 50 distinct images, and editing success is evaluated using a single VLM judge without human validation or statistical significance analysis.

Questions for authors

Please see weaknesses.

4 comments
AuthorsOpenReview ↗

W1. Validation of the copying mechanism & W2. Justification for the per-token objective

Dear Reviewer vFwB,

Thank you for your review and for engaging with our work! We are pleased to see that you found our paper novel, interesting, useful, and with strong empirical results. And that you noted we clearly demonstrated improved protection effectiveness and computational efficiency over existing baselines. Thank you for the additional comments. Please find our below our responses to your concerns.

Sincerely, The Authors

W1. Validation of the copying mechanism

We agree that our original analysis did not directly quantify reference copying. We conducted an additional evaluation measuring whole-image similarity between edited outputs and the original clean reference (not the perturbed image), aggregated across both benchmarks.

MethodDINO ↑R@1 (%) ↑LPIPS ↓SSIM ↑MS-SSIM ↑PSNR ↑
Unprotected.55185.3.458.528.46111.8
PhotoGuard.49981.6.489.546.50311.8
Diff-PGD.49576.2.621.329.4239.4
DeContext.22429.2.718.370.2329.4
CaD (ours).74091.2.418.535.70017.1

CaD achieves the strongest reference preservation, including substantially higher DINO similarity. Its 91.2% reference retrieval accuracy further confirms that outputs retain reference-specific information. Together with the controlled attention interventions in Fig. 2, these results support reference reconstruction rather than arbitrary generation failure.

Regarding low ArcFace similarity, we clarify that reference copying need not imply identity-faithful or pixel-perfect reconstruction. Our use of the term copying refers to the model's tendency to reproduce the reference image's content and spatial structure rather than carry out the requested edit. Moreoever, it can be observed in Figure 2 that sometimes, the attention output can become increasingly out of distribution, introducing some speckle and blur in the resulting image. This can affect the ArcFace similarity. We do not claim that the resulting reconstruction is lossless. CaD can preserve global scene structure and appearance while degrading fine-grained facial details, which disproportionately affects ArcFace. We will clarify our terminology as ``lossy reference reconstruction", explicitly distinguishing it from exact copying. We've also updated the paper to include this validation of the copying/reconstruction mechanism in the appendix section C. Please see the updated paper.

W2. Justification for the per-token objective

We agree that maximizing reference-attention mass does not guarantee coverage of individual identity-critical tokens. However, our token-coverage argument concerns the limitations of reference-attention suppression, rather than guaranteeing identity-critical token coverage in CaD.

Specifically, Fig. 1 illustrates why DeContext may succeed on face-centric images but fail on full-body images. In Fig. 1(a), the subject occupies 92% of reference tokens, making aggregate reference suppression more likely to disrupt subject information. In Fig. 1(b), the subject occupies only 31% (face: 4.8%), allowing the objective to potentially reduce reference attention substantially by suppressing background interactions without eliminating subject-related information.

But more importantly, token coverage interacts with attention redistribution. Since attention weights are softmax-normalized, suppressing reference attention redistributes probability mass toward text and target tokens. Thus, the editor may retain sufficient subject information while still attending to the pathways needed to execute the requested edit.

CaD addresses this limitation by reversing the direction of attention manipulation: maximizing reference attention simultaneously reduces attention available to text and target tokens. Our per-target-query objective encourages this redistribution across target queries, rather than only optimizing an aggregate image-level measure. It does not require or guarantee attention to specific identity-critical reference tokens.

We have updated the paper to clarify this as well as updated the token coverage claim about CaD.

AuthorsOpenReview ↗

W3. Motivation for attention manipulation versus alternative objectives & W4. Robustness Evaluation

W3. Motivation for attention manipulation versus alternative objectives

Our choice of attention manipulation is motivated by extensive prior work, controlled causal interventions, and computational efficiency.

First, attention is an established mechanism for controlling generative behavior in extensive prior work. Prior works such as Prompt-to-Prompt, MasaCtrl, and AdvPaint demonstrate its effectiveness for controlling image synthesis, preserving source content, and adversarial image protection. More importantly, our controlled interventions (Fig. 2) directly establish that increasing target-to-reference attention in early transformer blocks induces copying, whereas suppressing it does not reliably prevent editing. CaD exploits these observed behaviors.

Second, our existing baselines already span distinct optimization objectives: PhotoGuard targets VAE latent representations, Diff-PGD disrupts denoiser predictions, and DeContext suppresses reference attention. Thus, our evaluation compares CaD against both representation-level and denoiser-level alternatives, not exclusively attention-based defenses.

Third, attention manipulation offers a direct and computationally efficient objective. In in-context DiTs, increasing reference-attention mass necessarily reduces attention allocated to text and target tokens, directly limiting competing generation pathways. Feature-targeted losses do not explicitly enforce this redistribution, while direct output-level copying losses generally require differentiating through multiple denoising steps. In contrast, CaD backpropagates only through the early transformer blocks, requiring 0.647 s/step, versus 1.232 for DeContext and 1.245 for Diff-PGD on FLUX Kontext.

W4. Robustness Evaluation

Thank you for highlighting this concern. To address it, we have conducted additional robustness experiments evaluating cropping, Gaussian noise, and JPEG compression, with results presented below.

Consistent with prior work on adversarial image protection, we find that these transformations weaken protection effectiveness. Following purification, all methods exhibit broadly comparable performance across most metrics, although DeContext achieves lower ESR under cropping on both datasets. Notably, while DeContext achieves a lower ESR under cropping, its CLIP directional similarity and PickScore remain comparable to those of other methods. This suggests that editing may still occur, but in a decontextualized manner that is not counted as a successful edit by the VLM, rather than indicating consistently stronger robustness across metrics. These findings highlight a shared limitation of all existing protection methods. Improving robustness through techniques such as Expectation over Transformations (EoT) represents an important direction for future work.

We have updated the paper and included these additional robustness results and discussion in the appendix.

Table: Purification robustness at ε = 12/255 against FLUX Kontext. None denotes no purification. Lower is better for all metrics.

BodyBench

MetricMethodNoneCropGauss.JPEG
ESR ↓DeContext0.240.420.710.71
DiffPGD0.440.570.760.79
PhotoGuard0.810.760.780.82
Ours0.050.740.790.81
CLIPdir ↓DeContext0.270.260.270.28
DiffPGD0.250.270.300.30
PhotoGuard0.330.300.320.32
Ours0.030.290.310.32
PickScore ↓DeContext22.2921.6121.7821.78
DiffPGD21.3821.7322.0522.05
PhotoGuard22.2821.9322.2522.28
Ours18.6621.9322.1722.24

InpaintBench

MetricMethodNoneCropGauss.JPEG
ESR ↓DeContext0.210.410.640.60
DiffPGD0.520.480.730.68
PhotoGuard0.780.760.790.79
Ours0.160.760.780.80
CLIPdir ↓DeContext0.200.200.210.21
DiffPGD0.190.200.220.22
PhotoGuard0.240.220.230.24
Ours0.050.220.240.24
PickScore ↓DeContext21.8420.1920.5120.58
DiffPGD20.4820.4020.7020.72
PhotoGuard20.8920.7220.9320.84
Ours18.6520.6620.9120.88
AuthorsOpenReview ↗

W6. Other Works GuardDoor and PersGuard

W6. Other Works

Thank you for the references. However, these are fundamentally different deployment assumptions. Our work studies image-side proactive protection: an image owner constructs an adversarial perturbation and releases the protected image, without modifying the editing model or requiring the model provider to participate, which is a more realistic real-world scenario. This is the established threat model considered by the extensive body of work (including PhotoGuard, DeContext, etc.) and related adversarial image protection methods. But, the suggested works investigate a different setting:

  • GuardDoor (Zeng et al., 2025) requires cooperation from model providers to modify the image encoder and implant a protective backdoor.
  • PersGuard (Liu et al., 2026) similarly modifies pretrained model weights through protective backdoors We therefore do not consider model-side backdoor defenses directly comparable baselines for our setting. However, we agree that these works represent a complementary deployment scenario in which model providers actively participate in enforcing image protections.

We updated the paper and expanded our discussion to acknowledge GuardDoor and PersGuard as examples of provider-cooperative, model-side protection, while explicitly distinguishing this alternative deployment scenario from the image-side, owner-controlled protection studied in our work.

AuthorsOpenReview ↗

W7. Limited Evaluation Scale and Reliance on a Single VLM Judge

W7. Limited Evaluation Scale and Reliance on a Single VLM Judge

Thank you for raising this concern. We agree that human validation and statistical analysis strengthen the evaluation. However, our evaluation scale is consistent with established proactive image protection benchmarks, particularly given the computational cost of per-image optimization.

First, our scale is comparable to prior work. AdvPaint evaluates 100 images, DiffusionGuard introduces a benchmark with 42 base images, DiffVax evaluates 75 images, and DeContext uses 50 identities. Our evaluation spans 50 distinct images, 15 editing instructions per image, and two editing models, yielding 1,500 edited outputs per method, alongside multiple baselines and ablations. Each image requires separate optimization for every defense and model (approximately 16–20 minutes for several baselines), making larger-scale controlled comparisons computationally expensive.

Second, our benchmark selection specifically targets a gap in existing evaluations. Many prior benchmarks emphasize face-centric portraits, whereas our central finding concerns protection failures on full-body images, where identity-critical regions occupy substantially fewer reference tokens. We therefore combine face-centric images from InpaintGuardBench with full-body images from BodyBench to evaluate both settings.

Third, our evaluation does not rely exclusively on a single VLM judge. Alongside Qwen3-VL, we report PickScore, CLIP directional similarity, ArcFace similarity, and qualitative comparisons, consistent with established evaluation practices.

To further validate the automated judgments and address your concern, we conducted an independent human evaluation involving 76 participants and 500 edited outputs (50 source images × 2 editing models × 5 methods), with each output evaluated by at least three annotators. Participants viewed the original and edited images side by side, first assessing visual similarity, identity preservation, distortion, and usability without seeing the editing instruction, and then rating whether the requested edit was successfully performed on a five-point Likert scale. We report the percentage of ratings indicating agreement or strong agreement that the edit was performed, along with 95% bootstrap confidence intervals obtained by resampling source images.

Table 1: VLM ESR versus Human Editing Success

EditorMethodPaper VLM ESR (%) ↓Matched VLM ESR (%) ↓Human Edit Made (%) ↓ [95% CI]
FLUX KontextUnprotected81.886.093.3 [88.0, 97.3]
PhotoGuard80.380.086.0 [78.7, 92.7]
Diff-PGD47.350.074.7 [64.7, 84.0]
DeContext23.320.071.3 [62.0, 80.0]
Ours9.512.011.3 [4.0, 19.3]
FIBO EditUnprotected81.182.090.0 [83.3, 96.0]
PhotoGuard81.776.085.3 [78.0, 92.0]
Diff-PGD49.638.056.7 [45.3, 68.0]
DeContext63.764.077.3 [68.0, 86.0]
Ours41.248.050.0 [39.3, 60.7]

Table 1: Comparison of VLM-based and human-evaluated editing success rates. Paper VLM ESR is computed on the full benchmark, whereas Matched VLM ESR uses the subset selected for human evaluation. Human editing success measures whether the requested edit was performed, with 95% bootstrap confidence intervals. Each editor–method condition contains 50 edited outputs. Unlike the human metric, VLM ESR additionally requires identity preservation and the absence of major visual artifacts. Lower is better.

Importantly, the human evaluation reinforces our central motivation. For DeContext, there is a larger discrepancy between VLM ESR and human-judged instruction fulfillment. On FLUX Kontext, only 20.0% of outputs satisfy the VLM's stricter success criteria, yet humans identify the requested edit in 71.3% of outputs. This suggests that DeContext can still permit the requested transformation while disrupting the original image's identity or context, causing the VLM to classify these outputs as unsuccessful edits. This is precisely the limitation discussed in our paper: decontextualization does not necessarily prevent editing , motivating CaD's strategy of inducing copying or output corruption instead. In contrast, our method's VLM and human success rates closely agree and are consistently outperforming the other methods across models (12.0% vs. 11.3% on FLUX Kontext; 48.0% vs. 50.0% on FIBO Edit), providing additional support for its effectiveness.

We have updated the paper to include these additional experiments, incorporated the human evaluation results and uncertainty estimates into the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers (vFwB, HSBm, and QpsW) for their thoughtful feedback and recognition of CaD's strong contributions. We are encouraged that the reviewers highlighted the novelty and motivation of our method, strong empirical improvements over existing methods, computational efficiency, and prompt-agnostic protection.

To address the reviewers' concerns, we conducted four additional evaluations and updated the manuscript accordingly:

  1. Quantitative validation of copying: We measured whole-image similarity using DINO, R@1, LPIPS, SSIM, MS-SSIM, and PSNR. CaD achieves 0.74 DINO similarity and 91.2% reference retrieval accuracy, providing additional evidence of reference reconstruction on FLUX Kontext. This is now included in Table 2 in the revised manuscript.

  2. Robustness evaluation: We evaluated cropping, Gaussian noise, and JPEG compression across both benchmarks and all methods, identifying a shared limitation of existing adversarial protection approaches. This is now included in Table 3 in the revised manuscript.

  3. Cross-model transferability: We evaluated protection against three unseen editors without reoptimization. Notably, CaD reduces DreamOmni2's ESR from 58.4% to 0.9%, demonstrating strong transfer to a fine-tuned variant of FLUX Kontext. This is now included in Table 4 in the revised manuscript.

  4. Independent human evaluation: We conducted a human study, providing independent validation of CaD's effectiveness. Human judgments show that CaD substantially outperforms existing methods in preventing the requested edits, particularly on FLUX Kontext, where only 11.3% of CaD-protected outputs were judged successfully edited, compared with 71.3% for DeContext and 74.7% for Diff-PGD. This is now included in Table 5 in the revised manuscript.

We additionally updated the manuscript to addressed presentation suggestions. All new experimental results and corresponding discussions have been incorporated into the revised manuscript and appendix.

We believe these additions directly address the reviewers' principal concerns and strengthen the empirical and methodological support for CaD. We hope the reviewers can take these substantial revisions into account when making their final recommendation.

Thank you again!

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.