- 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
-
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.
-
The proposed method is cheaper and faster to optimize, which gives it good practical potential.
-
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
-
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.
-
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.
-
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.
-
Copying is only clearly shown on Flux Kontext. On FIBO, CaD outputs look like degraded images with blocky artifacts rather than copies (Fig. 3).
-
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
-
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.
-
The paper uses a lots phrasing such as "rather than ..." (10 times in the main text) and "not ... but ...", which hurts readability.
-
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
- How robust is CaD to JPEG compression, resizing and simple purification (e.g., noisy upscaling)?
- Does a perturbation optimized on Flux Kontext transfer to FIBO Edit, or to other editors?
- 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?
- How well does the Qwen3-VL judge agree with humans?
- How does the per-token objective compare with simply maximizing the average reference attention?
5 comments
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
| Editor | Method | Paper VLM ESR (%) ↓ | Matched VLM ESR (%) ↓ | Human Edit Made (%) ↓ [95% CI] |
|---|---|---|---|---|
| FLUX Kontext | Unprotected | 81.8 | 86.0 | 93.3 [88.0, 97.3] |
| PhotoGuard | 80.3 | 80.0 | 86.0 [78.7, 92.7] | |
| Diff-PGD | 47.3 | 50.0 | 74.7 [64.7, 84.0] | |
| DeContext | 23.3 | 20.0 | 71.3 [62.0, 80.0] | |
| Ours | 9.5 | 12.0 | 11.3 [4.0, 19.3] | |
| FIBO Edit | Unprotected | 81.1 | 82.0 | 90.0 [83.3, 96.0] |
| PhotoGuard | 81.7 | 76.0 | 85.3 [78.0, 92.0] | |
| Diff-PGD | 49.6 | 38.0 | 56.7 [45.3, 68.0] | |
| DeContext | 63.7 | 64.0 | 77.3 [68.0, 86.0] | |
| Ours | 41.2 | 48.0 | 50.0 [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.
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.
| Method | DINO ↑ | R@1 (%) ↑ | LPIPS ↓ | SSIM ↑ | MS-SSIM ↑ | PSNR ↑ |
|---|---|---|---|---|---|---|
| Unprotected | .551 | 85.3 | .458 | .528 | .461 | 11.8 |
| PhotoGuard | .499 | 81.6 | .489 | .546 | .503 | 11.8 |
| Diff-PGD | .495 | 76.2 | .621 | .329 | .423 | 9.4 |
| DeContext | .224 | 29.2 | .718 | .370 | .232 | 9.4 |
| CaD (ours) | .740 | 91.2 | .418 | .535 | .700 | 17.1 |
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.
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 against FLUX Kontext. None denotes no purification. Lower is better for all metrics.
BodyBench
InpaintBench
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.
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.