acceptodds
Under review as a conference paper at ICLR 2027

Key-Recoverable Conditional Flow Hijacking: Authorization-Aware Image Protection for DiT-Based In-Context Editing

Abstract

DiT-based in-context image editing has emerged as a powerful paradigm, but also enables unauthorized identity-preserving image manipulation, raising privacy and misuse concerns. Existing perturbation-based protections typically disrupt reference use indiscriminately, so legitimate users also lose the ability to perform reference-consistent editing once an image is protected. This motivates a more practical authorization-aware protection paradigm, moving beyond indiscriminate protection toward selective control over reference use. To this end, we introduce Key-Recoverable Conditional Flow Hijacking, an authorization-aware image protection framework for DiT-based in-context editing, where the same protected image blocks unauthorized reference-consistent editing while allowing authorized users to recover this capability with a valid credential. Our key insight is that the native visual- and text-conditioning pathways of a DiT act through different inputs but jointly influence the same target editing process, allowing a visual perturbation and a key embedding to form an asymmetric control pair. The visual perturbation disrupts unauthorized reference-conditioned editing, while the key embedding re-enables reference-consistent editing for authorized users. Our method only optimizes the protected image and its key embedding, leaving the pretrained editor unchanged and avoiding the cost of retraining or deploying an auxiliary recovery model. Experiments on FLUX.1-Kontext with CelebA-HQ and VGGFace2 demonstrate strong unauthorized protection and effective authorized recovery across standard and diverse editing prompts. For example, on CelebA-HQ under diverse edits, identity similarity drops from 0.78 for clean editing to 0.26 without the key and recovers to 0.68 with the correct key, while maintaining prompt alignment and visual quality. Further experiments verify key-specific recovery, sensitivity to protection hyperparameters, robustness to image transformations and purification, and transferability across unseen DiT-based editors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
DiT-based Image Editing, Authorization-Aware Protection, Conditional Flow Hijacking, Adversarial Perturbation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_j6XPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Key-Recoverable Conditional Flow Hijacking, which enables authorized users to perform reference-consistent editing on a protected image while preventing unauthorized users from obtaining reference-consistent edits. The key idea is to steer the editing trajectory so that the edited result of the perturbed image falls into the region that satisfies prompt consistency but not reference consistency. To do this, the authors optimize the perturbation by minimizing the hijacking loss between the context steering of the protected image and that of a hijack anchor, along with a repulsion loss from the clean context steering and an attention preservation loss. They further optimize a key embedding appended to the prompt embedding so that the edited results with the correct key fall into the region satisfying both the reference condition and the text condition. The perturbation and the key are optimized alternately while the pretrained editor is kept frozen. To demonstrate the effectiveness of the method, they conduct experiments on FLUX.1-Kontext with CelebA-HQ and VGGFace2, and further evaluate key selectivity, cross-model transferability, and robustness to image transformations and purification.

Critical strengths
  1. By introducing key parameters directly into the existing text path of the Diffusion Transformer (DiT), this method eliminates the need for additional modules or structural model modifications while generating distinct key values for each image.
  2. The underlying objectives are well designed through the steering decomposition in Equation 4, where the specific role of each loss component is thoroughly validated across transplantation, steering replacement, and ablation experiments.
  3. The approach demonstrates exceptional cross-model generalization where keys trained on FLUX.1-Kontext operate seamlessly across four other un-retrained editors, significantly boosting ICEdit's performance from 0.49 to 0.77.
Critical weaknesses
  1. Although the loss formulation seems appropriate, the core novelty of the loss formulation appears limited. The attention loss merely reverses the sign of the metric used in DeContext. Furthermore, LhijL_{\text{hij}} and LrepL_{\text{rep}} share the exact two-loss structure of Semantic Injection (Lee et al., 2026) which pulls toward the target semantics while suppressing the source semantics.
  2. The evaluation lacks robustness against adaptive attacks. Because the key operates as a continuous embedding space, an attacker could directly optimize a key to maximize identity similarity with the protected image. Merely testing against Random Key and Swap Key baseline scenarios fails to establish security against such tailored, white-box adaptive threats.
  3. The experimental scope and overall performance present notable weaknesses. The evaluation relies on a restricted dataset of only 50 face images per set, and selectivity is validated using just two prompts on CelebA-HQ. Even under Correct Key conditions, the protected outputs exhibit noticeable quality degradation compared to clean images, dropping in both ISM (from 0.78 to 0.68) and CLIP-IQA (from 0.81 to 0.70).
  4. Cross-model transfer experiments were mostly conducted on FLUX-family models, making it difficult to consider them true black-box transferability evaluations. Additionally, there is no explanation of how the key was applied to Step1X-Edit, which uses a different text encoder, nor is there a comparison with the transfer performance of baseline methods.
Questions for authors

Does protection and key restoration hold on Stable Diffusion based editing models like InstructPix2Pix or models with different architectures and text encoders like Qwen-Image-Edit? In particular, because UNet-based models utilize a cross-attention structure rather than joint attention, the proposed steering decomposition does not directly apply. Can you demonstrate whether the same image–key pairs remain functional in this setting, or if model-specific re-optimization is required?

5 comments
AuthorsOpenReview ↗

1、Response to Weakness 1 and Weakness 2

Response to Weakness 1: Novelty of the Loss Formulation

Reviewer’s Concern: Clarify the novelty of the proposed loss formulation and its technical differences from DeContext and Semantic Injection.

Response:

We thank the reviewer for the thoughtful analysis. We would like to clarify that our contribution is not a simple combination or sign reversal of existing losses, but a recoverable image editing protection mechanism based on Conditional Flow Hijacking, enabling unauthorized protection and key-based recovery without modifying the pretrained DiT or introducing additional auxiliary models.

Our protection mechanism fundamentally differs from DeContext. DeContext suppresses Target-to-Context Attention to weaken reference information propagation. As acknowledged in its Appendix D, this strategy becomes less effective in complex scene edits where generation is dominated by textual conditions and relies less on the visual reference. In contrast, our method preserves context interaction while actively redirecting reference-induced steering through the hijacking loss (LhijL_{hij}) and the representation loss (LrepL_{rep}). Thus, the attention loss (LattnL_{attn}) serves to maintain context engagement rather than directly enforce protection.

Our method also differs substantially from Semantic Injection (Lee et al., 2026), despite the similar attraction–repulsion formulation. Semantic Injection optimizes U-Net cross-attention outputs to inject predefined target semantics, providing one-way image protection. In contrast, our method redirects reference-induced Context Steering in DiT-based in-context editing to disrupt unauthorized identity preservation. More importantly, we jointly optimize an image-specific key embedding through the native text-conditioning pathway to restore reference-consistent editing. This asymmetric, key-recoverable control is not addressed by Semantic Injection.

The novelty of our method lies in integrating Context Steering Redirection, Context Engagement Preservation, and Key-conditioned Recovery to achieve image-level authorization without modifying the pretrained editor, rather than in the mathematical form of individual loss terms. We will clarify these distinctions in the revised manuscript.

Response to Weakness 2: Robustness Against Adaptive Attacks

Reviewer’s Concern: Random Key and Swap Key tests are insufficient to establish robustness against white-box adaptive attacks that optimize alternative key embeddings.

Response:

We thank the reviewer for this valuable suggestion. We agree that Random Key and Swap Key primarily evaluate key selectivity and do not fully assess robustness against adaptive attacks on continuous key embeddings. We therefore conducted an additional white-box adaptive key optimization experiment.

Specifically, we designed an attack in which the attacker has access to the publicly available protected image and the editing model, but not the clean image or the correct key. The attacker optimizes an alternative key embedding by maximizing ArcFace facial feature similarity between the edited output and the protected image, while keeping the editor frozen.

We evaluate the optimized key on CelebA-HQ using the fixed prompt "a photo of this person", with the following results:

MethodFDFRISM ↑CLIP-IQA ↑CLIP-Score ↑
Clean0.000.890.7824.13
Ours (No Key)0.000.140.7126.57
Ours (Optimized Key)0.020.210.6825.79
Ours (Correct Key)0.000.770.7024.37

The optimized alternative key increases ISM only slightly from 0.14 to 0.21, remaining well below the 0.77 achieved with the correct key. This indicates that the attacker fails to recover effective identity-preserving editing under the tested white-box attack setting.

We appreciate the reviewer for suggesting this meaningful attack scenario, which also opens up interesting directions for future research on adversarial attacks against multimodal models. Our results demonstrate robustness against the tested adaptive attack, and we will include this experiment and discussion in the revised manuscript.

AuthorsOpenReview ↗

2、Response to Weakness 3、Weakness 4 and Question [1/2]

Response to Weakness 3: Experimental Scope and Performance

Reviewer’s Concern: Limited evaluation scale and degraded authorized editing quality.

Response:

We thank the reviewer for raising concerns about evaluation scale, key selectivity, and authorized editing quality.

Regarding evaluation scale, we acknowledge that 50 reference images per dataset is relatively limited. Nevertheless, this is comparable to DeContext and NullEdit [1], which each evaluated 50 images per dataset, and CaD [2] and IDGuardian [3], which used 50 test images. Additionally, we evaluate 60 diverse prompts unseen during optimization, yielding 3,000 image-prompt combinations per dataset. Although broader identity coverage is desirable, evaluating 60 unseen prompts per reference provides substantial diversity in editing scenarios. We are currently expanding the datasets and sample sizes to further assess generalization.

Regarding Key Selectivity, we extend the original two-prompt evaluation to 60 diverse editing prompts, comparing No Key, Random Key, Swap Key, and Correct Key on CelebA-HQ.

SettingFDFRISM ↑CLIP-IQA ↑CLIP-Score ↑
Clean0.000.780.8126.56
No Key0.090.260.7627.67
Random Key0.090.260.7427.41
Swap Key0.000.280.7227.22
Correct Key0.000.680.7026.46

Across 60 diverse unseen prompts, Correct Key achieves an ISM of 0.68, substantially outperforming No Key (0.26), Random Key (0.26), and Swap Key (0.28). This confirms the image-specific selectivity of the learned keys.

Regarding authorized editing quality, we acknowledge the remaining gap from Clean (ISM: 0.78 vs. 0.68; CLIP-IQA: 0.81 vs. 0.70). Our key restores reference-consistent editing rather than removing the perturbation. Since authorized editing still uses the perturbed reference, residual identity deviations may contribute to the ISM gap, while perturbation artifacts may affect CLIP-IQA.

Appendix A.4.1 further demonstrates the protection–quality trade-off. Increasing the perturbation budget ε from 0.02 to 0.10 reduces No-Key ISM from 0.71 to 0.14, while Correct-Key ISM decreases from 0.97 to 0.77 and CLIP-IQA from 0.75 to 0.70. A smaller budget can improve authorized editing quality but may weaken protection.

We are investigating ways to improve recovery and reduce visual artifacts while maintaining protection.

Response to Weakness 4 and Question: Cross-Model Transferability and Key Adaptation [1/2]

Reviewer’s Concern: The cross-model transfer evaluation lacks comparisons with existing protection baselines.

1. Comparison with Baseline Protection Transfer

We additionally evaluate AdvDM and DeContext on Step1X-Edit and ICEdit. Since these baselines provide protection without key-based recovery, we compare their No-Key transfer performance.

Target EditorMethodFDFRISM ↓CLIP-IQA ↑CLIP-Score ↑
Step1X-EditClean0.000.830.7524.36
AdvDM0.000.500.6024.06
DeContext0.000.600.5924.34
Ours (No Key)0.000.420.5423.73
ICEditClean0.000.910.7823.46
AdvDM0.000.820.5724.10
DeContext0.000.690.6423.70
Ours (No Key)0.000.490.4923.57

Our method achieves stronger identity disruption on both editors. On Step1X-Edit, ISM is 0.42, compared with 0.50 for AdvDM and 0.60 for DeContext. On ICEdit, it is 0.49, compared with 0.82 and 0.69, respectively. We compare all methods on the same editors using the same identity metric. Since AdvDM and DeContext lack key-based recovery, this comparison concerns unauthorized protection only.

Unlike these protection-only baselines, our method also supports recovery: Correct Key increases ISM from 0.42 to 0.63 on Step1X-Edit and from 0.49 to 0.77 on ICEdit, without target-specific re-optimization.

Reviewer’s Concern: Explanation of how the key was applied to Step1X-Edit, which uses a different text encoder.

2. Key Injection Across Different Text Encoders and Architectures

For Step1X-Edit, we truncate the FLUX.1-Kontext key from 16 × 4096 to 16 × 3584, then append it as additional tokens to the encoded text-conditioning sequence.

We use the same strategy for Qwen-Image-Edit and InstructPix2Pix: the original FLUX key is truncated to match the target text-conditioning dimension and appended to its encoded text embeddings. This reuses the original key without additional trainable adapters or target-editor modifications. The same dimension-matching truncation is applied to the other editors.

[1] Huang et al. NullEdit. arXiv:2608.10870, 2026.

[2] Abdel Magid et al. CaD. arXiv:2610.35566, 2026.

[3] Xiong et al. IDGuardian. CVPR, 2026.

AuthorsOpenReview ↗

3、Response to Weakness 4 and Question [2/2]

Response to Weakness 4 and Question: Cross-Model Transferability and Key Adaptation [2/2]

Reviewer’s Question: Does protection and key recovery transfer across different architectures, such as Qwen-Image-Edit and InstructPix2Pix, or is model-specific re-optimization required?

Reponses3-4:

3. Protection and Authorized Recovery Across Different Architectures

We further evaluate transferability on Qwen-Image-Edit and the U-Net-based InstructPix2Pix. All protected images and keys are optimized on FLUX.1-Kontext and reused directly, without target-specific re-optimization.

Target EditorSettingFDFR ↓ISM ↑CLIP-IQA ↑CLIP-Score ↑
Qwen-Image-EditClean0.000.670.7323.90
No Key0.000.310.7124.66
Correct Key0.450.500.6522.25
InstructPix2PixClean0.020.970.7823.53
No Key0.160.710.4623.27
Correct Key0.180.840.6723.43

These results show partial protection and recovery transfer, but also limitations across heterogeneous architectures. On Qwen-Image-Edit, No-Key ISM decreases from 0.67 to 0.31, with CLIP-IQA remaining close to Clean. The Correct Key raises ISM to 0.50 among face-detected outputs; however, FDFR reaches 0.45, meaning 45% of outputs contain no detectable face and recovery remains unstable.

On InstructPix2Pix, No-Key ISM decreases from 0.97 to 0.71, and the Correct Key raises it to 0.84. However, No-Key CLIP-IQA drops to 0.46; the Correct Key improves it to 0.67, but it remains below Clean.

We acknowledge that our Context Steering decomposition, formulated for DiT joint attention, does not directly apply to the cross-attention structure of U-Net models. Differences in visual conditioning, text representations, and denoising dynamics also make cross-architecture transfer challenging. Thus, while protective perturbations show partial transfer, directly reusing the FLUX key does not guarantee stable, high-quality recovery.

4. Discussion on Cross-Model Authorization and Potential Key Adaptation

We regret that our method still has limitations in cross-model generalization, particularly across editing architectures and text-conditioning interfaces that differ substantially from those evaluated in our experiments.

We further clarify the role of cross-model transfer in the server-side authorization scenario illustrated in Fig. 1. In our framework, image owners publicly release protected images, while authorized users hold access credentials. The Protection Service Provider verifies user authorization and manages the keys associated with each protected image. From this perspective, cross-model transferability is a desirable capability rather than a strict requirement of our authorization mechanism. The essential requirement is to verify whether a user is authorized to edit a protected image, rather than to ensure that a single fixed key embedding is universally compatible with every editor. If different editors require different key embeddings, the Protection Service Provider can associate model-specific keys with the same user credential and provide the appropriate key for the selected editor. In this way, authorized users can access different editing models using their corresponding keys, without republishing the protected image or changing the overall authorization workflow.

Motivated by this consideration, we have further thought about a potential direction for reducing the cost of cross-model adaptation. One possible approach is Warm-Started Cross-Model Key Adaptation, which initializes the target-model key from the existing FLUX key and learns only a model-specific residual:

Ek(m)=Adapt(EkFLUX)+ΔEk(m).E_k^{(m)} = \mathrm{Adapt}(E_k^{\mathrm{FLUX}}) + \Delta E_k^{(m)}.

Here, Adapt(⋅)\mathrm{Adapt}(\cdot) adjusts the original key to the target conditioning dimension, while ΔEk(m)\Delta E_k^{(m)} is a residual embedding optimized for the target editor. The protected image and pretrained target editor remain frozen, and only the residual key is optimized. Specifically, the target editor's predictions conditioned on the clean reference image can serve as supervision to align the editing dynamics of the protected image with those of clean editing. In this way, the existing FLUX key provides a warm start, potentially reducing the adaptation effort compared with optimizing a new key from random initialization.

We emphasize that this is a preliminary idea rather than an established solution, and its effectiveness across different architectures and text encoders remains to be investigated. We will add the results once they become available.

We hope that the above clarifications and additional experiments address the reviewer's concerns. We will incorporate the corresponding explanations, experimental results, and discussions into the revised manuscript.

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

To mitigate potential privacy risks arising from unauthorized editing and misuse of facial images, this paper proposes an authorization-aware image protection framework for image editing. The authors leverage the visual- and text-conditioning pathways of DiTs to establish an asymmetric control mechanism using visual perturbations and key embeddings. The perturbations disrupt unauthorized reference-consistent editing, while the corresponding key embeddings enable authorized editing. The method only optimizes the protected image and its associated key embedding without modifying the pretrained editor or requiring an auxiliary recovery model.

Critical strengths
  1. Elegant Method Design. By alternately optimizing visual perturbations and key embeddings, the authors establish an asymmetric control mechanism that disrupts reference-consistent editing for unauthorized users while enabling authorized users to achieve reference-consistent editing with the correct key. This novel and straightforward design effectively balances image protection and authorized editing.

  2. Comprehensive Experimental Evaluation. The paper presents extensive comparative and ablation studies to validate the effectiveness of the proposed method. The experimental results demonstrate substantial performance improvements over existing approaches, achieving state-of-the-art results and highlighting the effectiveness and competitiveness of the proposed method.

Critical weaknesses
  1. Insufficiently Highlighted Contributions and Limited Interpretability. The paper does not clearly highlight the key contributions of the proposed method compared with existing approaches. Furthermore, the interaction mechanism between visual perturbations and key embeddings is not sufficiently analyzed or theoretically explained, limiting the interpretability of the proposed framework.

  2. Lack of Comparisons with Recent Methods. The paper lacks comparisons with recently proposed state-of-the-art methods, which limits the comprehensiveness of the experimental evaluation and makes it difficult to fully assess the advantages of the proposed approach over the latest methods.

3.Insufficient Explanation of Ablation Metrics. Although the paper provides ablation studies, the evaluation metrics and their relevance to the proposed method are not sufficiently explained, making it difficult to fully interpret the experimental results and assess the contributions of individual components.

Questions for authors
  1. Could the authors provide additional comparisons with relevant state-of-the-art methods proposed in 2026 to better demonstrate the competitiveness of the proposed approach?
  2. Could the authors further clarify the core contributions and novelty of the proposed method, particularly its key technical differences and advantages over existing authorization-aware image protection methods?
  3. The CLIP-IQA scores reported in the ablation study are relatively low. Could the authors explain the possible reasons for these results and discuss whether the perceptual quality can be improved without compromising protection effectiveness and authorized editing performance?
3 comments
AuthorsOpenReview ↗

1、Response to Weakness 1 and Question 2: Core Contributions and Mechanistic Interpretability

Response to Weakness 1 and Question 2: Core Contributions and Mechanistic Interpretability

Reviewer’s Concern: Clarify the core contributions and novelty of the proposed method, and explain the interaction mechanism between visual perturbations and key embeddings.

Response:

We thank the reviewer for highlighting the need to clarify our contributions and the interaction between visual perturbations and key embeddings.

Limitations of existing methods. Existing authorization-aware protection methods, such as Adapter Shield, primarily target U-Net-based image-to-image generation and rely on separately trained encryption/decryption modules for authorized recovery. They do not address authorization-aware protection in DiT-based in-context editing. Meanwhile, DeContext targets DiT-based in-context editing but protects images by suppressing target-to-context attention, thereby weakening the influence of visual context on generation. As acknowledged in its Appendix D, this strategy becomes less effective in complex, text-driven editing scenarios, where the model already pays limited attention to visual context and further attention suppression has limited impact.

Our core contribution is a cross-pathway asymmetric authorization mechanism specifically designed for DiT-based in-context editing. It has two key technical components. First, Conditional Flow Hijacking preserves target-to-context attention but redirects reference-dependent context steering, rather than simply suppressing the reference's influence. Second, an image-specific learnable Key Embedding enables authorized recovery through the editor's native text-conditioning pathway, without modifying pretrained model weights or introducing additional recovery networks. Together, these components provide protection against unauthorized editing while allowing the corresponding key to recover identity-preserving generation.

Interaction between visual perturbations and key embeddings. Our steering decomposition shows that target-side generation dynamics are influenced by three components: text steering, target self-steering, and context steering. The visual perturbation primarily acts through the reference-context pathway, redirecting context steering to disrupt unauthorized identity preservation. In contrast, the corresponding Key Embedding acts through the text-conditioning pathway, adjusting text steering to compensate for the perturbation-induced deviation in target-side generation dynamics. Importantly, the key does not remove the visual perturbation; authorized recovery is achieved by conditioning the generation process on the protected image and its corresponding key.

This mechanism is supported by our Functional Disentanglement of Conditional Steering analysis in Sec. 4. Through counterfactual steering transplantation, we show that context steering predominantly governs reference-dependent information, whereas text steering primarily governs prompt-dependent editing semantics. These observations provide empirical evidence for the distinct yet complementary roles of the visual and text-conditioning pathways.

In summary, our novelty lies in combining a context-steering-based protection mechanism with native text-conditioned authorization for DiT-based in-context editing. The steering decomposition and counterfactual intervention experiments further provide a mechanistic interpretation of how the protected image and its corresponding key jointly influence generation.

AuthorsOpenReview ↗

2、Response to Weakness 2 and Question 1: Comparisons with Recent Methods

Response to Weakness 2 and Question 1: Comparisons with Recent Methods

Reviewer’s Request: Provide additional comparisons with relevant state-of-the-art methods proposed in 2026 to better demonstrate the competitiveness of the proposed approach

Response:

We thank the reviewer for this suggestion. We surveyed recent work on authorization-aware image protection, which remains relatively underexplored in image editing. Adapter Shield [1] is among the closest approaches to our objective, but we could not find a publicly accessible official implementation for reproducible comparison.

For DiT-based in-context editing protection, we identified DeContext [2], VETO [3], and CaD [4]. DeContext is already included in our experiments, while publicly accessible official implementations of VETO and CaD could not be found.

To broaden our evaluation, we additionally reproduced two recent image-protection methods, AnonymizerNoise [5] and IDGuardian [6], on FLUX.1-Kontext using CelebA-HQ and the standard prompt "a photo of this person."

MethodFDFRISM ↓CLIP-IQA ↑CLIP-Score ↑
Clean0.000.890.7824.13
AnonymizerNoise (2026)0.000.830.7224.59
IDGuardian (2026)0.000.790.6824.27
Ours (No Key)0.000.140.7126.57

Our method achieves substantially lower identity similarity (ISM = 0.14) while maintaining comparable visual quality and higher prompt alignment. Since these two methods were not originally designed for DiT-based in-context editing, this experiment evaluates their applicability to our setting rather than providing a matched white-box comparison.

Importantly, our method also supports authorized editing: the correct image-specific key restores ISM to 0.77, without image reconstruction, editor modification, or an auxiliary recovery network.

References

[1] Jia, Jun, et al. "Adapter Shield: A Unified Framework with Built-in Authentication for Preventing Unauthorized Zero-Shot Image-to-Image Generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 30120–30129.

[2] Shen, Linghui, Mingyue Cui, and Xingyi Yang. "DeContext as Defense: Safe Image Editing in Diffusion Transformers." arXiv preprint arXiv:2512.16625 (2025).

[3] Grebe, Jonas, et al. "VETO: Towards Protecting Images From Frontier AI Editing." arXiv preprint arXiv:2607.27292 (2026).

[4] Abdel Magid, Salma, et al. "Copying as Defense: Protecting Images from Editing by Eliciting Model Copying." arXiv preprint arXiv:2610.35566 (2026).

[5] Seo, Sanghyeok, Jiwoo Kim, and Jong-Uk Hou. "Anonymization-Guided Latent Perturbation for Source-Identity Reduction with Limited Edit Deviation in Instruction-Guided Image Editing." Sensors 26, no. 19 (2026): 6003.

[6]Xiong, Lizhi, et al. "No Way To Steal My Face: Proactive Defense Against Identity-Preserving Personalized Generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 20680–20690.

AuthorsOpenReview ↗

3、Response to Weakness 3 and Question 3

Response to Weakness 3: Clarification of Ablation Metrics

Reviewer’s Concern: Clarify the ablation metrics and their relevance to evaluating the contributions of individual loss components.

Response:

We thank the reviewer for this comment. Our ablation study examines how each loss term contributes to unauthorized identity protection and authorized identity recovery. Accordingly, the metrics should be interpreted under both authorization conditions.

ISM is the primary metric for evaluating these objectives. Under No Key, lower ISM indicates stronger disruption of reference identity; under Correct Key, higher ISM approaching the clean baseline indicates more effective recovery. FDFR measures face detection failure, while CLIP-IQA and CLIP-Score evaluate visual quality and prompt alignment, respectively.

With these objectives in mind, Table 4 reveals the distinct contributions of the three loss terms. Lhij\mathcal{L}_{\mathrm{hij}} provides an explicit target for redirecting context steering toward the hijack anchor. Removing it increases No-Key ISM from 0.26 to 0.73, approaching the clean baseline of 0.78, highlighting its critical role in identity disruption. Lrep\mathcal{L}_{\mathrm{rep}} separates protected context steering from clean-reference steering. Without it, ISM rises to 0.47, indicating that this separation further strengthens unauthorized protection. Lattn\mathcal{L}_{\mathrm{attn}} preserves target-to-context attention, encouraging steering redirection rather than context suppression. Removing it increases ISM to 0.42, demonstrating the benefit of maintaining context engagement.

Meanwhile, Correct-Key ISM remains between 0.68 and 0.73 across all ablations, with FDFR consistently at 0.00. This suggests that these loss terms primarily strengthen unauthorized protection while leaving authorized recovery relatively stable. The full method achieves the largest separation between No-Key and Correct-Key identity similarity (0.26 vs. 0.68), demonstrating their complementary contributions to asymmetric authorization.

We will clarify the metric interpretations, their directions under different authorization conditions, and their connections to the ablation results in the revised manuscript.

Response to Question 3: Perceptual Quality and Protection Trade-off

Reviewer’s Question: Explain the relatively low Correct-Key CLIP-IQA scores and whether perceptual quality can be improved without compromising protection effectiveness and authorized editing performance.

Response:

We thank the reviewer for this question. We believe the relatively low CLIP-IQA scores reflect a trade-off between protection strength and perceptual quality.

Our visual observations show that, although the Correct Key restores reference identity, some perturbation patterns from the protected image may carry over into the edited output, introducing visual artifacts. This is because the key restores reference-consistent editing rather than explicitly removing protective perturbations. Consequently, stronger perturbations can improve unauthorized protection while potentially reducing the perceptual quality of authorized edits, explaining why Correct-Key CLIP-IQA does not fully recover to the Clean level.

The perturbation-budget analysis in Appendix A.4.1 further supports this trade-off. Under the standard editing prompt, increasing ϵ\epsilon from 0.02 to 0.10 reduces No-Key ISM from 0.71 to 0.14, indicating stronger protection, while Correct-Key CLIP-IQA decreases from 0.75 to 0.70. Reducing the perturbation budget can therefore improve perceptual quality, but may weaken unauthorized protection.

These findings suggest that the perturbation budget can be adjusted to balance protection effectiveness and editing quality. A promising direction is to suppress perturbation artifacts in authorized outputs, for example through perceptual quality constraints, while preserving unauthorized protection and authorized recovery. We will clarify this trade-off in the revised manuscript.

We hope that the above clarifications and additional experiments address the reviewer's concerns. We will incorporate the corresponding explanations, experimental results, and discussions into the revised manuscript.

Reviewer_28RBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies image protection that disrupts unauthorized reference-consistent editing while allowing recovery through an image-specific embedding key. With a frozen DiT editor, it optimizes a bounded image perturbation to redirect context-attention contributions and a text-side key to recover clean-reference target updates. Experiments use FLUX.1-Kontext, two face datasets, and standard and held-out editing prompts. On diverse CelebA-HQ edits, identity similarity changes from 0.78 for clean inputs to 0.26 without the key and 0.68 with the correct key. Additional experiments examine mismatched keys, component ablations, transfer, and post-processing.

Critical strengths

1.Demonstrates selective protection and recovery without modifying the editor. On diverse CelebA-HQ edits, identity similarity decreases from 0.78 for clean editing to 0.26 without the key and recovers to 0.68 with the correct key. This establishes a useful protection/recovery effect on the same publicly released image.

2.Provides intervention-based mechanism evidence. Exchanging clean and protected context steering substantially changes identity similarity, while removing the hijacking loss increases unauthorized similarity from 0.26 to 0.73. These experiments support the functional importance of the targeted representation and loss.

Critical weaknesses

1.The proposed mechanism’s incremental contribution is insufficiently isolated. DeContext already protects images through reference-context manipulation, and Adapter Shield already supports authorized recovery (Sec. 2). The critical missing baseline is DeContext protection followed by the identical text-side key optimization. Since the recovery loss does not directly constrain the image update, current comparisons do not establish whether the proposed hijacking mechanism provides a distinctive protection/recovery advantage (Algorithm 1; Tables 1, 4).

2.The authorization boundary is insufficiently validated across independent models and adaptive reuse. Several evaluated editors share FLUX-based foundations, including ACE++ and ICEdit and DreamOmni2, limiting architectural diversity. Additional testing on independently developed models such as Qwen-Image-Edit would strengthen the transfer claim. Moreover, random/swapped-key failures do not establish resistance to alternative embeddings, prompt search, or preprocessing followed by model switching; purification already partially restores identity similarity (Tables 2, 5, 11). Such adaptive evaluation is missing.

Questions for authors

1.Incremental contribution: How does the proposed method compare with DeContext-protected images paired with the identical text-side key-recovery procedure, under matched optimization budgets, key capacity, and perceptual distortion? This comparison would clarify whether the hijacking objective provides a distinctive advantage for recoverable protection beyond combining existing protection with learned conditioning (Algorithm 1; Tables 1, 4).

2.Generalization and adaptive robustness: Can the unchanged protected images prevent identity-preserving editing on independently developed models such as Qwen-Image-Edit, including after budgeted preprocessing, purification, or prompt search? Please distinguish unauthorized protection transfer from authorized-key recovery, and report thresholded identity-preserving edit success alongside similarity and detection failures. These results would clarify the authorization boundary beyond the current model-transfer and mismatched-key tests (Sec. 3.1; Tables 2, 5, 11).

5 comments
AuthorsOpenReview ↗

1、Response to Weakness 1 and Question 1: Incremental Contribution

Response to Weakness 1 and Question 1: Incremental Contribution

Reviewer’s Concern: The reviewer questions the novelty of our method over DeContext and Adapter Shield, and asks whether the proposed hijacking mechanism provides a distinct advantage over DeContext combined with the same key-recovery procedure under matched settings.

Response:

We thank the reviewer for raising this important point. We apologize that the original manuscript did not sufficiently clarify the distinction between our approach and existing protection or authorization methods. We appreciate the opportunity to further clarify our technical contributions and the novelty of our approach.

First, Adapter Shield relies on separately trained encryptor and decryptor modules for authorized recovery, whereas our method requires neither additional encoder nor decoder modules. Instead, we introduce an image-specific learnable soft key through the editor's native text-conditioning pathway, without modifying pretrained model or training additional recovery modules. Moreover, Adapter Shield does not address authorization-aware protection for DiT-based in-context image editing; it focuses on zero-shot image-to-image generation using U-Net-based diffusion models, whereas we specifically target DiT-based in-context editing. Thus, our contribution extends authorization-aware image protection to a distinct editing paradigm rather than merely adopting an existing recovery mechanism.

Second, our protection mechanism is fundamentally different from DeContext. DeContext protects images by suppressing target-to-context attention, thereby weakening the influence of the reference image on generation. However, as acknowledged in Appendix D of the DeContext paper, this strategy becomes less effective under complex, text-driven scene modifications, where the model already attends weakly to the visual context and further attention suppression has limited impact. Our Conditional Flow Hijacking instead maintains target-to-context attention while redirecting reference-dependent context steering. Rather than further suppressing context attention, we change how reference information guides the editing process, providing a different protection mechanism.

Our central contribution is therefore not simply combining image protection with a learned key, but designing a context-steering-based protection mechanism that supports key-conditioned recovery through the native conditioning pathway of DiT-based in-context editors.

To isolate this contribution, we implement the reviewer's suggested DeContext + Key baseline. Specifically, we apply the identical text-side key optimization procedure to DeContext-protected images, matching the protection and key optimization budgets, key capacity, and perturbation constraints. We evaluate both methods on CelebA-HQ using 60 diverse editing prompts.

MethodNo-Key FDFRNo-Key ISM ↓No-Key CLIP-IQA ↑No-Key CLIP-Score ↑Correct-Key FDFRCorrect-Key ISM ↑Correct-Key CLIP-IQA ↑Correct-Key CLIP-Score ↑
Clean0.000.780.8126.56————
DeContext + Key0.010.570.7627.240.000.670.7026.43
Ours0.090.260.7627.670.000.680.7026.46

Without the key, our method achieves substantially stronger protection than DeContext + Key (ISM: 0.26 vs. 0.57), with identical CLIP-IQA (0.76) and comparable prompt alignment. With the correct key, both methods achieve nearly identical recovery quality (ISM: 0.68 vs. 0.67; CLIP-IQA: 0.70 for both). Notably, the No-Key-to-Correct-Key ISM gap is 0.42 for our method, compared with only 0.10 for DeContext + Key. Under matched settings, these results demonstrate that Conditional Flow Hijacking provides a distinctive protection advantage while preserving comparable authorized recovery, beyond simply adding text-side key optimization to an existing defense.

Finally, the recovery loss does not update the image perturbation by design. We first optimize the image perturbation to establish protection independently of the key, then freeze the protected image and optimize the key for authorized recovery. This separation ensures that recovery is achieved through the conditioning key rather than by modifying the protected image to accommodate it. Allowing the recovery objective to update the perturbation could introduce a competing optimization pressure that compromises no-key protection. Our matched DeContext + Key experiment further shows that, even with the same recovery procedure, the image-side protection objective remains critical to achieving a strong authorization gap.

AuthorsOpenReview ↗

2、Response to Weakness 2 and Question 2: Cross-Model Generalization and Adaptive Robustness [1/2]

Response to Weakness 2 and Question 2: Cross-Model Generalization and Adaptive Robustness [1/2]

We thank the reviewer for these concerns. Beyond the original Random Key and Swap Key tests, we evaluate an independent editor, alternative-key optimization, prompt search, and preprocessing followed by model switching.

Reviewer’s Request: Report thresholded identity-preserving success.

1. Thresholded Identity-Preserving Success.

We introduce Protection Success Rate (PSR) alongside ISM and FDFR. Using DeepFace's default ArcFace threshold of 0.68 cosine distance, we set the corresponding cosine similarity threshold to 0.32. [email protected] measures the proportion of outputs with no detected face or similarity below 0.32; higher values indicate stronger protection.

Reviewer’s suggestion: Additional testing on independently developed models such as Qwen-Image-Edit would strengthen the transfer claim.

2. Cross-Model Transfer to Qwen-Image-Edit. We directly transfer FLUX.1-Kontext-optimized images and keys to Qwen-Image-Edit without model-specific re-optimization.

SettingFDFRISM ↓[email protected] ↑CLIP-IQA ↑CLIP-Score ↑
Clean0.000.672.04%0.7323.90
Ours (No Key)0.000.3165.31%0.7124.66
Ours (Correct Key)0.450.5054.37%0.6522.25

Without the key, ISM decreases from 0.67 to 0.31 and PSR reaches 65.31%, while FDFR remains 0.00, demonstrating protection transfer without re-optimization. The Correct Key raises the average ISM of face-detected outputs from 0.31 to 0.50. However, FDFR reaches 0.45, meaning 45% of outputs contain no detectable face. Thus, the original key provides partial identity recovery, but authorized recovery remains unstable across this architecture.

Reviewer’s Concern: Random/Swap Key tests may be insufficient to establish robustness against adaptive attacks, including alternative-embedding optimization, prompt search, and preprocessing/purification followed by model switching.

Response (3-5):

3. Adaptive Attack Using Alternative Embeddings. We conduct a white-box attack where the attacker accesses the protected image and frozen editor, but not the clean reference or correct key. The attacker optimizes an alternative embedding to maximize ArcFace similarity between the output and protected image. We perform 200 optimize steps and evaluate on CelebA-HQ using "a photo of this person".

SettingFDFRISM ↑[email protected] ↓CLIP-IQA ↑CLIP-Score ↑
Clean0.000.896.12%0.7824.13
Ours (No Key)0.000.1497.96%0.7126.57
Ours (Optimized Key)0.020.2187.50%0.6825.79
Ours (Correct Key)0.000.7710.20%0.7024.37

The alternative key increases ISM from 0.14 to 0.21, whereas the Correct Key achieves 0.77. PSR remains 87.50% with the optimized key, compared with 10.20% with the Correct Key, indicating limited recovery under the tested attack budget.

4. Adaptive Attack Through Prompt Search. We adapt RATTPO's [1] test-time prompt optimization, using ArcFace similarity between the output and protected image as the reward. Without a key, the attacker iteratively refines prompts based on previous prompts and rewards (e.g., from "A person sitting down" to "Generate a seated person while retaining their original facial features"). Due to computational cost, we independently optimize 50 image-prompt pairs (10 images × 5 instructions).

SettingFDFRISM ↑[email protected] ↓CLIP-IQA ↑CLIP-Score ↑
Protected0.060.2572.30%0.786727.9884
Prompt Search0.020.3169.40%0.797529.4831

Prompt Search increases ISM from 0.25 to 0.31, while PSR remains at 69.40%, indicating that protection remains effective for a substantial proportion of outputs.

5. Preprocessing Followed by Model Switching. We apply JPEG preprocessing to protected images before editing them with Step1X-Edit, which was not involved in the original optimization.

SettingFDFRISM ↓[email protected] ↑CLIP-IQA ↑CLIP-Score ↑
Clean (JPEG + Step1X)0.000.832.04%0.7524.36
No Key (JPEG + Step1X)0.000.4150.37%0.5324.15
Correct Key (JPEG + Step1X)0.000.6513.17%0.6724.17

After JPEG preprocessing and transfer to Step1X-Edit, PSR remains at 50.37% without the key. With the Correct Key, ISM increases from 0.41 to 0.65 and PSR decreases to 13.17%, demonstrating retained recovery after preprocessing and model switching.

[1] Kim, Semin, et al. "Reward-agnostic prompt optimization for text-to-image diffusion models." arXiv preprint arXiv:2506.16853 (2025).

AuthorsOpenReview ↗

3、Response to Weakness 2 and Question 2: Cross-Model Generalization and Adaptive Robustness [2/2]

Response to Weakness 2 and Question 2: Cross-Model Generalization and Adaptive Robustness [2/2]

6. Discussion on Cross-Model Authorization and Potential Key Adaptation

We regret that our method still has limitations in cross-model generalization, particularly across editing architectures and text-conditioning interfaces that differ substantially from those evaluated in our experiments.

We further clarify the role of cross-model transfer in the server-side authorization scenario illustrated in Fig. 1. In our framework, image owners publicly release protected images, while authorized users hold access credentials. The Protection Service Provider verifies user authorization and manages the keys associated with each protected image. From this perspective, cross-model transferability is a desirable capability rather than a strict requirement of our authorization mechanism. The essential requirement is to verify whether a user is authorized to edit a protected image, rather than to ensure that a single fixed key embedding is universally compatible with every editor. If different editors require different key embeddings, the Protection Service Provider can associate model-specific keys with the same user credential and provide the appropriate key for the selected editor. In this way, authorized users can access different editing models using their corresponding keys, without republishing the protected image or changing the overall authorization workflow.

Motivated by this consideration, we have further thought about a potential direction for reducing the cost of cross-model adaptation. One possible approach is Warm-Started Cross-Model Key Adaptation, which initializes the target-model key from the existing FLUX key and learns only a model-specific residual:

Ek(m)=Adapt(EkFLUX)+ΔEk(m).E_k^{(m)} = \mathrm{Adapt}(E_k^{\mathrm{FLUX}}) + \Delta E_k^{(m)}.

Here, Adapt(⋅)\mathrm{Adapt}(\cdot) adjusts the original key to the target conditioning dimension, while ΔEk(m)\Delta E_k^{(m)} is a residual embedding optimized for the target editor. The protected image and pretrained target editor remain frozen, and only the residual key is optimized. Specifically, the target editor's predictions conditioned on the clean reference image can serve as supervision to align the editing dynamics of the protected image with those of clean editing. In this way, the existing FLUX key provides a warm start, potentially reducing the adaptation effort compared with optimizing a new key from random initialization.

We emphasize that this is a preliminary idea rather than an established solution, and its effectiveness across different architectures and text encoders remains to be investigated. We will add the results once they become available.

We hope that the above clarifications and additional experiments address the reviewer's concerns. We will incorporate the corresponding explanations, experimental results, and discussions into the revised manuscript.

Comments on the paper

1 comment

Dear Reviewer Mkst,

Thank you for taking the time to consider our responses and additional experiments. We appreciate your thoughtful feedback and are glad that our clarifications have addressed several of your concerns. We will incorporate the additional results and corresponding discussions into the revised manuscript.

Thank you again for your time and careful consideration of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.