- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes DeContext, an adversarial perturbation-based defense against unauthorized identity-preserving image editing in Diffusion Transformer (DiT) models.
The authors first conduct attention intervention experiments and identify the attention connections from target queries to context keys as an important pathway for preserving identity information from reference images. Motivated by this observation, they formulate an adversarial optimization objective that minimizes the attention mass assigned by target tokens to context tokens. The input image is then optimized using Projected Gradient Ascent (PGA) under a bounded perturbation constraint, while keeping the model parameters frozen. Furthermore, the authors analyze the importance of different diffusion timesteps and Transformer blocks, concentrating the optimization on early denoising stages and selected single-stream blocks.
Experiments on VGGFace2 and CelebA-HQ, using FLUX Kontext and Step1X-Edit, demonstrate that DeContext can substantially reduce identity similarity between the reference and generated images while maintaining relatively natural-looking outputs.
Critical strengths
S1. Insightful mechanism analysis and a simple yet effective method. The authors investigate the influence of different attention directions between target and context tokens through controlled attention intervention experiments. Their findings highlight the importance of target-query-to-context-key attention connections in preserving identity information from reference images. This observation provides useful insights into contextual information propagation in DiT-based image editing models and offers a clear motivation for the proposed method. Building on this finding, the authors formulate context attention suppression as a constrained optimization problem. By applying Projected Gradient Ascent to the input image while keeping model parameters frozen, DeContext reduces the attention allocated to context information. The method is conceptually simple, and its effectiveness is further improved by focusing on influential diffusion timesteps and Transformer blocks. The clear connection between the initial mechanism analysis and the subsequent method design is a key strength supporting the contribution of this work.
S2. Promising experimental results supporting the effectiveness of the proposed method. The authors evaluate DeContext on VGGFace2 and CelebA-HQ and demonstrate its effectiveness on two DiT-based image editing models, FLUX Kontext and Step1X-Edit. For example, on CelebA-HQ with the prompt "A photo of this person," DeContext reduces the identity similarity score (ISM) from 0.79 to 0.12, substantially weakening identity preservation while producing relatively natural-looking images. In addition, the ablation studies on perturbation budgets, diffusion timesteps, and Transformer blocks provide empirical support for the main design choices. The authors also conduct preliminary robustness evaluations under common image preprocessing operations. These results provide convincing evidence of the method's effectiveness under the evaluated settings and constitute an important strength of the paper.
Critical weaknesses
W1. Insufficient fine-grained quantitative evidence for the causal relationship between attention connections and identity preservation. The authors demonstrate the importance of target-query-to-context-key attention through the attention masking experiments in Figure 2. They further analyze attention distributions and attack effectiveness across diffusion timesteps and Transformer blocks in Figure 4 and Table 5. However, these experiments do not directly quantify the causal contribution of specific attention connections to identity preservation. In particular, a higher attention mass does not necessarily imply a greater contribution to information transfer or to the final generated output. Therefore, while the experiments support the effectiveness of the proposed intervention strategy, the underlying explanation of contextual information propagation remains somewhat incomplete. This limitation weakens the paper's contribution in terms of understanding the internal mechanisms of DiT-based image editing models.
W2. Insufficient evaluation of adversarial perturbation generalization across models and prompts. As a defense intended for practical image protection, the effectiveness of DeContext depends not only on its performance under specific optimization settings but also on whether the resulting perturbations remain effective under unseen generation conditions. First, although the authors evaluate DeContext on Step1X-Edit, the current description does not clearly establish whether the same adversarial image optimized for one model can be directly transferred to another model without re-optimization. Since real-world adversaries may freely choose different image editing models, this issue is important for assessing the practical effectiveness of the defense. Second, the main quantitative evaluations of facial identity protection cover a relatively limited number of prompts. Moreover, some evaluation prompts in Table 2 overlap with the prompt pool used during perturbation optimization. Consequently, the current experiments provide limited evidence for generalization to diverse and previously unseen editing instructions. These limitations reduce my confidence in the robustness and reliability of DeContext under open-world attack scenarios.
Questions for authors
Q1. On the causal relationship between attention connections and identity preservation. The authors provide relevant evidence through attention masking, gradient analysis, and block-wise ablation studies. However, to what extent do these experiments establish the direct causal contribution of specific attention connections to identity preservation? Could the authors conduct more fine-grained interventions on individual Transformer layers or attention heads, and quantitatively analyze the relationship between intervention strength, changes in attention mass, and the resulting identity similarity? Such evidence would help clarify whether attention mass serves as a reliable indicator of contextual information contribution and would strengthen the proposed mechanism explanation. The results could materially affect my assessment of the paper's interpretability contribution.
Q2. On cross-model transferability of adversarial perturbations. Could the authors clarify the experimental setup in Section 4.4? Specifically, does the Step1X-Edit evaluation directly use adversarial perturbations optimized for FLUX Kontext, or are the perturbations re-optimized specifically for Step1X-Edit? If re-optimization is required, can the same protected image remain effective when directly transferred to another image editing model without any additional optimization? I would particularly appreciate cross-model transfer results between FLUX Kontext and Step1X-Edit. Such experiments would distinguish the applicability of the proposed optimization method across architectures from the transferability of the resulting adversarial perturbations themselves. This distinction is important for assessing the practical value of the proposed defense.
Q3. On prompt generalization. The authors evaluate two prompts excluded from the optimization prompt pool in Table 1 and provide additional results under different editing instructions in the main paper and appendix. However, the main quantitative evaluations still cover a relatively limited set of prompts. Furthermore, some evaluation prompts in Table 2 overlap with the optimization prompt pool in Appendix F.3. For example, "Add glasses to this person." appears in both. Could the authors clarify the distinction between seen and unseen prompts in the current evaluation and provide results on a broader, independent set of test prompts? In particular, it would be valuable to separately evaluate unseen prompts that are semantically similar to the optimization instructions and prompts belonging to editing categories not covered by the optimization pool. Such results would help establish the extent to which DeContext generalizes to previously unseen editing scenarios.