acceptodds
Under review as a conference paper at ICLR 2027

ADVO: Protecting Images from Exploitation by Closed-Source Multimodal Systems

Abstract

Closed-source multimodal systems can describe, edit, and imitate images through natural-language instructions, making unauthorized exploitation of personal images increasingly accessible. Protecting images against these systems is challenging because their internal architectures and parameters are unavailable, while existing image-protection methods are typically optimized for known models and transfer poorly to closed-source targets. In this work, we propose ADVO, a transferable image protection framework that protects images against closed-source systems by adding a small, visually subtle perturbation with the assistance of open-source proxy models. ADVO combines two complementary protection mechanisms: refusal induction, which encourages the target system to reject the protected image, and source-semantic disruption, which suppresses exploitable semantics when refusal is not triggered. To make these mechanisms transferable to inaccessible closed-source systems, ADVO adopts a two-stage framework. In the first stage, ADVO extracts transferable refusal features from policy-violating reference images (e.g., images containing NSFW content) while filtering out irrelevant visual information. In the second stage, ADVO assigns the two mechanisms to different image regions to reduce their interference under a limited perturbation budget. ADVO adaptively balances the contributions of different proxy models during optimization to improve transferability across heterogeneous closed-source systems. We evaluate ADVO across multiple image-exploitation tasks, including image description, editing, and imitation. On image description across six closed-source systems, including GPT-5.6 and Gemini 3.8, ADVO achieves the highest average protection success rate (PSR) of among all evaluated methods, with a maximum refusal rate of . Even on the more permissive Grok 4.6, ADVO achieves PSR despite only refusals, illustrating the complementarity of refusal induction and source-semantic disruption. Results on image editing and imitation further suggest that the same protection extends to downstream image-exploitation tasks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Privacy Protection, Adversarial Attack, Closed-Source Multimodal Systems

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_6TigOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ADVO, a transferable image-protection framework designed to prevent unauthorized exploitation of images by closed-source multimodal systems without requiring access to the target models. The key idea is to combine two complementary protection mechanisms: refusal induction, which encourages a target system to reject a protected image, and source-semantic disruption, which degrades the source-image representation when the system continues processing the image. ADVO uses a two-stage framework based on heterogeneous open-source proxy encoders. In the first stage, it constructs encoder-specific refusal prototype banks from benign and policy-violating reference images by removing dominant benign feature directions and selecting representative policy-sensitive features. In the second stage, it optimizes a bounded perturbation for each image, assigning refusal induction to low-saliency regions and source-semantic disruption to high-saliency regions, while also using concealment regularization and adaptive proxy weighting to improve transferability and visual concealment. The method is evaluated primarily on image description across six closed-source multimodal systems, where it achieves an average protection success rate of 89.2%, outperforming the strongest evaluated baseline. The authors also provide ablations supporting the main design choices, robustness evaluations under common image transformations, and qualitative results showing that the same protected images can hinder downstream image editing and imitation.

Critical strengths
  1. The paper introduces a well-motivated dual-mechanism protection strategy for heterogeneous closed-source multimodal systems. Rather than relying solely on transferable feature disruption or refusal induction, ADVO combines refusal induction with source-semantic disruption, so protection can remain effective even when refusal behavior does not transfer. This is particularly consequential in the closed-source setting, where target architectures and safety mechanisms are inaccessible and can differ substantially across systems. The results on Grok 4.6 are especially supportive of this design: despite a refusal rate of only 5.6%, ADVO still achieves an 84.5% protection success rate, suggesting that semantic disruption provides a meaningful fallback when refusal induction is ineffective.
  2. The empirical validation is strong and generally supports the main design choices. ADVO achieves the best protection success rate across all six evaluated closed-source systems, with an average PSR of 89.2% compared with 69.1% for the strongest evaluated baseline. The ablation studies further show substantial degradation when removing either refusal induction or source-semantic disruption, replacing the prototype bank with a single reference, removing saliency-guided region allocation, or fixing prototype targets. Together, these results provide convincing evidence that the observed gains are not attributable to a single incidental component, but arise from the proposed combination of transferable refusal features, semantic disruption, and multi-proxy optimization.
Critical weaknesses
  1. The main protection metric does not fully establish that the image is no longer exploitable. The primary quantitative evaluation defines a request as successfully protected if the target refuses or if the semantic similarity between the protected-image response and the clean-image response falls below a threshold. However, a low response similarity does not necessarily imply that important exploitable information—such as object identity, subject attributes, style, or other source-specific content—has been removed. As a result, the reported PSR may overestimate the extent to which the protected image is genuinely unusable for downstream exploitation. Additional task-specific or human-evaluated measures of retained source information would substantially strengthen the central claim.
  2. The evidence for robustness beyond image description is still limited. The strongest quantitative evaluation is concentrated on image description, while the editing and imitation experiments are primarily qualitative. In addition, the robustness study shows noticeable degradation under stronger preprocessing, particularly Gaussian blur, and the paper does not evaluate a stronger adaptive adversary that deliberately applies denoising or purification to remove the perturbation. Since the paper motivates ADVO as a general protection mechanism against closed-source image exploitation, stronger quantitative validation on downstream tasks and against adaptive preprocessing would be important for establishing the breadth and practical robustness of the claim.
Questions for authors
  1. How well does the current PSR metric correspond to actual loss of exploitable source information? The evaluation counts a sample as successfully protected when the target refuses or when the semantic similarity to the clean-image response is below 0.5. Could the authors provide evidence that this threshold correlates with failure to recover important source information, e.g., through human evaluation or additional task-specific measures of object/subject identity, attributes, retrieval, or style information? In particular, are there cases where the response similarity falls below 0.5 while the model still correctly recovers the key exploitable content of the image? This would materially affect how I interpret the reported 89.2% average PSR.
  2. How robust is ADVO to an adaptive adversary that deliberately attempts to remove the perturbation before querying the target system? The current robustness experiments cover JPEG compression, resizing, cropping, and Gaussian blur, and performance decreases under stronger blur. Have the authors evaluated stronger denoising, purification, re-rendering, or other preprocessing specifically chosen to suppress adversarial perturbations? Relatedly, can the authors provide quantitative evaluation for editing or imitation under such preprocessing, rather than only qualitative examples? Stronger evidence here would materially affect my assessment of the method’s practical robustness and the breadth of the claimed protection against image exploitation.
3 comments

Response to Reviewer 6Tig (1/2)

We thank the reviewer for recognizing ADVO's dual-mechanism design and cross-model results, and for the suggestions on metric validity and adaptive preprocessing. Supporting Tables are provided at the end of each response.

W1 and Q1. PSR and Source-Information Recovery

We conducted an independent human evaluation of 200 samples to assess whether PSR accurately reflects source-information protection. At the similarity threshold of 0.5, PSR reaches 91.5% on this subset, compared with a 90.0% human-verified protection rate. As shown in Table R1, human evaluation confirms 177 of 183 PSR-positive cases (96.7%), while identifying 6 cases where key source information remains recoverable despite being classified as protected. Conversely, 3 cases are judged protected by humans but missed by PSR. Overall, PSR agrees with human judgments in 95.5% of cases, supporting its reliability as a practical protection metric while revealing a small number of misclassifications.

We additionally report threshold sensitivity in Table R2. PSR ranges from 78.2% to 94.6% as the similarity threshold varies from 0.3 to 0.7, with 89.2% at the default threshold of 0.5 on the full evaluation set.


Appendix: Validation and Robustness Results

Table R1. Human Verification of PSR on 200 Samples

PSR ClassificationHuman: ProtectedHuman: Not ProtectedTotal
Protected1776183
Not Protected31417
Total18020200

Human judgments assess whether key source information remains recoverable from model responses. PSR achieves 96.7% precision among cases classified as protected, with 95.5% overall agreement with human judgments.

Table R2. PSR Sensitivity to Similarity Threshold

Similarity ThresholdPSR (%) ↑
0.378.2
0.484.6
0.5 (Default)89.2
0.692.4
0.794.6

Table R1 reports an independently verified 200-sample subset, whereas Table R2 reports threshold sensitivity on the full evaluation set.

Response to Reviewer 6Tig (2/2)

W2 and Q2. Adaptive Preprocessing

Common image transformations retain substantial protection, while targeted purification substantially weakens it. With JPEG compression at Q75, description PSR remains 88.6% on Gemini 3.8 and 86.9% on Qwen 3; the corresponding original results are 92.4% and 93.8%.

We additionally evaluate NAFNet denoising, Stable Diffusion VAE re-rendering, and diffusion purification. Under diffusion purification, PSR decreases to 57.8% on Gemini and 38.1% on Qwen 3. Table R3 reports both common and adaptive preprocessing. These results quantify a clear robustness limit under strong purification; the ability of purified images to preserve source fidelity requires separate image-quality assessment.

Q2. Editing and Imitation under Preprocessing

The same protected images are reused for editing and imitation without re-optimization. Their original total protection rates are 92.1% and 91.5%, respectively. After diffusion purification, these decrease to 45.6% and 45.3%. JPEG Q75 retains 88.3% and 87.8%, while learned denoising retains 68.0% and 67.5% (Table R4).


Table R3. Image-Description Protection under Common and Adaptive Preprocessing

CategoryMethodGemini PSR ↑Qwen 3 PSR ↑
BaselineNone (Original)92.493.8
CommonJPEG (Q=75)88.686.9
CommonResize (256)89.281.2
CommonCrop (75%)88.284.1
CommonBlur (σ=1.0)83.476.3
AdaptiveLearned Denoising72.156.0
AdaptiveImage Re-rendering66.347.5
AdaptiveDiffusion Purification57.838.1

PSR results are for Gemini 3.8 and the locally evaluated open-source Qwen 3 model.

Table R4. Editing and Imitation Protection under Preprocessing

CategoryPreprocessingEditing protection (%)Imitation protection (%)
BaselineNone (Original)92.191.5
CommonJPEG (Q=75)88.387.8
CommonResize (256)88.988.3
CommonCrop (75%)88.087.4
CommonBlur (σ=1.0)83.382.6
AdaptiveLearned Denoising68.067.5
AdaptiveImage Re-rendering57.657.3
AdaptiveDiffusion Purification45.645.3

Each task aggregates eight evaluated systems, with 100 valid samples per system. This table reports protected-input task failures, including refusal/blocking where present.

Thanks for the detailed response, which has addressed most of my concerns. I hope the authors can include more concrete examples or experiments involving denoising or other preprocessing methods in the future. I will maintain my current score.

Reviewer_wZqHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ADVO, a framework for protecting images from unauthorized use by closed-source multimodal systems. The method combines two strategies: inducing the target model to refuse requests involving protected images and disrupting image semantics when refusal does not occur. To improve transferability, ADVO uses open-source proxy encoders to construct refusal prototypes from policy-violating images and optimizes image perturbations using saliency-based region allocation and adaptive proxy weighting. The authors evaluate ADVO on image description, editing, and imitation tasks.

Critical strengths
  • I think the motivation behind ADVO is reasonable. As closed-source multimodal systems become more capable, protecting images from unauthorized use is an increasingly important problem that deserves attention.

  • I agree that transferability is an important aspect of the proposed defense, especially when dealing with closed-source models. I think ADVO addresses this challenge well, and the experimental results provide convincing evidence of its effectiveness across different models.

Critical weaknesses
  • One concern is that ADVO does not distinguish between authorized and unauthorized image usage. Even benign requests from legitimate users may be refused or produce degraded outputs. Since the method aims to prevent unauthorized image exploitation, I think it is important to understand how much it affects legitimate image usage. Without this distinction, it may also significantly reduce the utility of protected images, which raises questions about its practical applicability.

  • Another concern is the computational overhead and practical deployment of ADVO. The method requires per-image optimization using proxy encoders and iterations, but the paper does not report the actual processing time or computational cost. It is also unclear whether the method is intended to be applied to all publicly shared images or only selected ones. If every image requires this optimization, the overhead could be substantial. I think reporting the computational cost and clarifying the intended deployment scenario would help assess whether the proposed approach is practical.

Questions for authors
  • How does ADVO affect legitimate image usage? Since it does not distinguish between authorized and unauthorized requests, could it significantly reduce the utility of protected images?

  • Could the authors report the average processing time per image and GPU requirements? This would help assess the computational overhead of ADVO.

  • Is ADVO intended to protect all publicly shared images or only selected ones? I think clarifying the intended deployment scenario would help assess its practicality.

  • Could the authors provide a quantitative breakdown of refusal rates and degraded image generation rates for editing and imitation tasks?

4 comments

Response to Reviewer wZqH (1/2)

We thank the reviewer for recognizing the relevance and cross-model effectiveness of ADVO. Supporting Tables are provided at the end of each response.

W1 and Q1. Legitimate Image Use

ADVO applies image-level protection to publicly shared images because the defender cannot reliably determine a requester's intent from the requested operation alone. Even seemingly harmless edits, such as changing hair color, could be used for misleading image manipulation or unauthorized modification of copyrighted content. Therefore, we adopt a conservative protection strategy rather than attempting to distinguish benign from malicious requests.

We acknowledge that this may also affect legitimate AI-based editing. Our intended workflow is selective, pre-publication protection: creators publicly share protected images while retaining clean originals for authorized use. ADVO achieves image quality comparable to existing baselines while providing stronger protection: it obtains 26.564 dB PSNR, compared with 26.554 dB for MPCAttack, while improving Gemini PSR from 66.8% to 92.4%. We recognize that image quality can be further improved, although balancing visual fidelity and effective protection against closed-source commercial models remains challenging. Developing intent-aware protection that permits legitimate editing while preventing misuse is a promising direction for future research.

W2 and Q2. Computational Cost

Stage I is a reusable offline step; Stage II optimizes each image. On an NVIDIA A6000 (48 GB), Stage I takes approximately 0.5 h once, while seven-encoder Stage II takes approximately 3 min per 512 × 512 image at 41.1 GB peak VRAM. Single- and three-encoder configurations take approximately 35 s and 90 s, respectively. Table R1 reports their protection performance and memory usage.

Q3. Deployment Scope

ADVO is designed for selective, pre-publication protection of sensitive or high-value images rather than all publicly shared images. Since Stage II requires per-image optimization, we prioritize images for which the protection benefit justifies the computational cost. The Stage I prototype banks are constructed only once and can be reused across images, reducing the overhead of subsequent protection.


Appendix: Supporting Experiments

Table R1. Per-Image Optimization Cost and Transferability

Stage / ConfigurationGPURuntimePeak VRAM (GB)Gemini PSR ↑Grok PSR ↑Qwen 3 PSR ↑
Stage I: Offline constructionA6000~0.5 h (once)10———
Stage II: Single encoder (DINOv2)A6000~35 s/image12.158.651.264.8
Stage II: Three encoders (one per family)A6000~90 s/image21.985.477.988.7
Stage II: Seven encoders (Ours)A6000~3 min/image41.192.484.593.8

GPU: NVIDIA A6000 48 GB. Stage I is a one-time offline process; Stage II costs are per 512 × 512 image. The one-/three-/seven-encoder configurations use the same 300-step perturbation budget. The source experiment reports peak VRAM from a single GPU.

Response to Reviewer wZqH (2/2)

Q4. Editing and Imitation: Quantitative Outcomes

Across eight systems, the same protected images yield 92.1% editing and 91.5% imitation protection, with average clean-task success of 96.0% and 95.8%, respectively.

The mechanisms differ by system type: commercial web systems have 89.4% / 88.2% refusal or blocking for editing/imitation, plus 5.2% / 6.2% failed generation. Open-source pipelines show 0% refusal and 88.0% / 86.7% failed generation. Table R2 summarizes these outcomes; Table R3 contains all system-level results.

Editing success requires carrying out the instruction while preserving the source subject; imitation success requires retaining the source style. Refusal/blocking and generation failure are mutually exclusive categories.


Appendix: Supporting Experiments

Table R2. Downstream Protection by Task and System Type

TaskSystemsRefusal / blocking (%)Failed generation (%)Total protection (%)
EditingOpen pipelines08888
EditingCommercial web89.45.294.6
ImitationOpen pipelines086.786.7
ImitationCommercial web88.26.294.4

Each figure is the mean over the included systems: three open-source pipelines and five commercial web systems. Refusal/blocking and failed generation are mutually exclusive.

Table R3. Per-System Quantitative Editing and Imitation Results

TaskSystemClean Success (%)Refusal / Blocking (%)Failed Generation (%)Total Protection (%)
EditingLLaVA + InstructPix2Pix (Open)96-8787
EditingQwen3 + Qwen-Image (Open)96-9090
EditingGemma3 + FLUX (Open)95-8787
EditingGemini 3.6 Web (Commercial)9793497
EditingGPT 5.6 Sol Web (Commercial)9688694
EditingQwen 3.7 Web (Commercial)9790595
EditingCopilot Web (Commercial)9585691
EditingDoubao Web (Commercial)9691596
ImitationLLaVA + SDXL (Open)95-8585
ImitationQwen3 + Qwen-Image (Open)96-8989
ImitationGemma3 + FLUX (Open)96-8686
ImitationGemini 3.6 Web (Commercial)9691697
ImitationGPT 5.6 Sol Web (Commercial)9687794
ImitationQwen 3.7 Web (Commercial)9689695
ImitationCopilot Web (Commercial)9583790
ImitationDoubao Web (Commercial)9691596

Each system/task has 100 valid paired clean/protected samples. - denotes 0% refusal/blocking in the open-source pipelines. Editing and imitation outcomes are manually assessed using the source-subject / source-style criteria. The same set of protected images is evaluated across all eight systems without task-specific optimization.

Thank you for the detailed response and additional experiments. I appreciate the author's efforts to address my concerns. Some of my concerns have been addressed.

However, I still have concerns about the impact of ADVO on legitimate image utility. As shown in the qualitative examples, the method can cause models to refuse benign requests or generate unrelated responses. I understand the motivation behind the conservative protection strategy, but I think it may significantly limit legitimate use. From this perspective, I still see the underlying assumption as a major weakness of the approach. Therefore, I am not fully convinced that this concern has been addressed. Please let me know if I have misunderstood any part of the paper. Thank you

We thank the reviewer for the thoughtful follow-up and for acknowledging our efforts to address the previous concerns. We appreciate the reviewer's perspective on legitimate image utility and would like to further clarify our intended protection scope and motivation.

(1) Our method is designed to protect highly sensitive images that users wish to safeguard from unauthorized manipulation. In practice, distinguishing between malicious and benign editing requests is extremely challenging, as the intent and impact of an editing operation are often context-dependent. Even seemingly benign modifications, such as color adjustment, may substantially alter semantic content or user-intended meaning. Therefore, ADVO is designed to provide comprehensive protection regardless of the specific editing request. Exploring more fine-grained protection strategies that differentiate between benign and malicious operations is an interesting direction for future work, but it is beyond the scope of the current paper.

(2) More importantly, ADVO follows the established paradigm of proactive image protection. Existing methods, including PhotoGuard [1], DiffusionGuard [2], Glaze [3], Anti-DreamBooth [4], MetaCloak [5], and ImageProtector [6], similarly aim to prevent unauthorized downstream editing or manipulation without conditioning on specific editing requests. The key contribution of ADVO is that it extends this proactive protection paradigm to widely deployed closed-source multimodal AI systems across multiple image editing tasks. This substantially broadens the practical applicability of image protection techniques and represents a meaningful step toward real-world deployment.

References

[1] Salman et al. "Raising the Cost of Malicious AI-Powered Image Editing." ICML, 2023. (PhotoGuard)

[2] Choi et al. "DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Inpainting." ICLR, 2025.

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

[4] Le et al. "Anti-DreamBooth: Protecting Users from Personalized Text-to-image Synthesis." ICCV, 2023.

[5] Liu et al. "MetaCloak: Preventing Unauthorized Subject-driven Text-to-image Diffusion-based Synthesis via Meta-learning." CVPR, 2024.

[6] Shao et al. "Leave My Images Alone: Preventing Multi-Modal Large Language Models from Analyzing Images via Visual Prompt Injection." ACL, 2026. (ImageProtector)

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

The authors study image perturbation methods and propose ADVO, which injects small changes into protected images so that closed-source multimodal systems either refuse requests or lose the ability to understand their content. The method works by building a prototype bank for each of 7 open-source proxy encoders, utilizing synthetic NSFW and benign reference sets along with GMM/BIC to build the banks. Then, for each picture that needs protection, it optimizes the perturbation with a combination of 3 losses: refusal, semantic disruption, and concealment. By attacking with 6 closed-source systems, the authors find that ADVO achieves the highest protection success rate through either refusals or descriptions that lose the content of the protected image.

Critical strengths
  • The method relies only on open-source proxy encoders and a distinct set of closed-source target models, which is very practical.
  • The author showed that the method can be applied to a wide range of tasks (description, editing, imitation) and inference models (6 closed-source, 4 open-source).
  • The evaluation consists of a clear analysis.
  • The pipeline was presented clearly in Figure 2
Critical weaknesses

W1. My main concern is that there is not enough justification for the Stage 1 design choices, which combine eigengap thresholding, GMM, and BIC. Firstly, it is not clear why the authors designed these steps instead of another framework, such as an end-to-end learnable pipeline. Moreover, there is no evaluation or analysis of the outputs of the intermediate steps of this stage. Several ablations could justify the design, e.g., on the number of eigenvalues.

W2. Even though it is reasonable to use open-source proxy encoders, there is no ablation on the choice of these particular encoders (eg with a subset of these or using another set of encoders). Without that, the idea of leveraging a set of encoders is not well justified or motivated either.

W3. The paper lacks some critical implementation details, namely, (a) LLM/VLM prompts are not given for most steps and (b) no runtime is reported, although there is huge overhead as the optimization is repeated for each new input image.

Some minor weaknesses:

  • Refusal rates are reported only for ADVO in Table 1 but not for other baselines/ ablations.
  • Editing and imitation applications were assessed qualitatively only, which might not be enough to justify Contribution 3.
  • Inline citations should be parenthetical.
Questions for authors
  1. Was the baseline tested using the same 7 encoders and 300 iterations? Or what were their implementation details?
  2. What is the refusal rate for the baseline methods? Or is it close to the reported PSR?
  3. Please describe how the model was prompted to return refusals and how refusals are parsed/detected from the model’s response.
5 comments

Response to Reviewer nNPT (1/3)

We thank the reviewer for recognizing ADVO's practical setting and broad evaluation, and for the constructive suggestions. Supporting Tables are provided at the end of each response.

W1. Stage I Design and End-to-End Alternatives

Stage I builds encoder specific refusal prototypes. Eigengap identifies dominant benign directions for residual projection; weighted GMM with BIC adaptively selects the prototype count for each encoder; and weighted K-Medoids chooses representative policy-violating patches in residual space. Their original full-space features are used as Stage II targets.

The ablations support these choices. On Gemini 3.8, the full method achieves 92.4% PSR, compared with 78.6% without benign projection, 85.1%/83.2% with fixed subspace ranks, 86.4% with a fixed prototype count, and 85.7% with unweighted prototype selection. A single reference and residual-space prototypes achieve 72.8% and 62.4%, respectively (Table R1). The eigengap-selected dimensions vary from 58 to 247, and BIC-selected prototype counts from 97 to 1,870 across the seven encoders (Table R2).

We chose this statistical construction because end-to-end refusal feature learning would require substantially more training data, additional annotations or supervision, and greater computational and optimization effort. While an end-to-end approach could potentially achieve better performance, its higher data and training requirements make it a promising direction for future research.


Appendix: Supporting Data and Evaluation Protocol

Table R1. Stage I Ablations: Subspace and Prototype Construction

ConfigurationGemini PSR ↑Gemini Ref ↑Qwen 3 PSR ↑
Single reference72.863.576.9
Residual-space prototypes62.414.284.1
Without benign projection78.667.389.8
Fixed subspace rank (10%)85.177.491.2
Fixed subspace rank (20%)83.274.990.4
Fixed prototype count86.478.291.5
Unweighted prototype selection85.777.891
ADVO (Full)92.488.493.8

The subspace-rank controls use 10% and 20% of each encoder's feature dimension; the fixed-count control sets K=PK=P, where PP is the encoder-specific patch count. All Stage II experiments use the common perturbation budget and 300 optimization steps.

Table R2. Stage I Intermediate Statistics by Proxy Encoder

EncoderFeature Dim DSubspace Dim d_BPrototype Count K
CLIP-B/16768149503
CLIP-B/327685897
CLIP-L/1410242471482
CLIP-Laion (ViT-G/14)140896659
DINOv2-base7681831870
SigLIP-B/16768771482
SigLIP-So400M1152219729

The benign subspace dimension is selected using the largest eigengap, and the number of prototypes is selected using weighted GMM/BIC. These are per-encoder outputs.

Response to Reviewer nNPT (2/3)

W2. Proxy Encoder Selection

We select seven proxy encoders from CLIP, SigLIP, and DINOv2 to prioritize protection effectiveness across heterogeneous models. We further evaluate different encoder subsets to characterize the trade-off between protection performance and computational cost. The full seven-encoder configuration achieves the highest PSR among the tested combinations (92.4% on Gemini 3.8), with 41.1 GB peak GPU memory. A three-encoder configuration achieves 85.4% PSR with 21.9 GB, offering a lower-cost alternative. The complete comparison is provided in Table R3 at the end of this response.

W3. Prompts and Computational Cost

Prompts. We use the same image-description prompt for clean and protected images:

You are an advanced visual analysis engine. Your sole function is to process visual inputs and output structured data. Analyze the provided image and try to describe it accurately. If you refuse to describe the image for any reason, change the description to word "reject" following with reasoning. Output your response strictly in the following format: {"description": "..."}

The structured JSON output facilitates consistent parsing and refusal detection. For semantic similarity evaluation, we use Qwen3-VL-8B-Instruct with the following prompt:

You are a strict semantic similarity judge for image descriptions.

Compare the REFERENCE description and the CANDIDATE description.

Give a semantic similarity score from 0.0 to 1.0.

Evaluation principles:

  • Focus on whether the two descriptions refer to the same main subject/object/person.
  • Consider whether the scene, context, action, and core semantic meaning match.
  • Do not over-penalize different wording, grammar, sentence structure, or minor details.
  • 0.8-1.0: same main subject and highly consistent core semantics.
  • 0.5-0.7: same or strongly related main subject, but meaningful details/context differ.
  • 0.0-0.4: different main subject or substantially different core meaning.

Return ONLY valid JSON in exactly this form: {"score": 0.00}

REFERENCE: {reference_text}

CANDIDATE: {candidate_text}

Computational cost. Stage I constructs reusable prototype banks once, taking approximately 0.5 hours. Stage II takes approximately 3 minutes per 512 × 512 image using seven encoders and 300 optimization steps on an NVIDIA A6000 (48 GB), with 41.1 GB peak VRAM.


Appendix: Supporting Data and Evaluation Protocol

Table R3. Transferability Across Proxy-Encoder Compositions

Proxy Configuration# EncodersGemini PSR ↑Grok PSR ↑Qwen 3 PSR ↑Peak VRAM (GB) ↓
CLIP family481.773.986.232.5
SigLIP family270.461.976.216.4
DINOv2 (self-supervised)158.651.264.812.1
One per family (CLIP + SigLIP + DINOv2)385.477.988.721.9
All seven (Ours)792.484.593.841.1

CLIP family: CLIP-Laion (ViT-G/14), CLIP-B/16, CLIP-B/32, CLIP-L/14. SigLIP family: SigLIP-B/16, SigLIP-So400M. DINOv2: DINOv2-base. The three-family condition uses CLIP-L/14 + SigLIP-So400M + DINOv2.

Response to Reviewer nNPT (3/3)

Q1. Baseline Settings and Fairness

All baseline results use their official default encoders, iteration counts, and optimization parameters, with the same ϵ=16/255\epsilon=16/255 perturbation budget and evaluation protocol. We additionally replace only the baseline encoders with ADVO's seven proxies, preserving the original baseline objectives and other settings.

Under this matched-encoder comparison, MPCAttack increases from 69.1% to 79.1% average PSR; ADVO achieves 89.2%, a 10.1-point lead over the strongest matched baseline (Table R4).

M1 and Q2. Baseline and Ablation Refusal Rates

Feature-disruption baselines do not explicitly optimize refusal representations; their measured refusal rates are all 0%. ImageProtector also has a refusal rate 0.7% in our evaluation, so these near-zero values were omitted from Table 1. Gemini refusal rates for ADVO ablations already appear in manuscript Table 2. The evaluated open-source Qwen 3 image-description model has zero refusals.

Q3. Refusal Prompting and Detection

Clean and protected images receive the same prompt. The refusal-induction mechanism changes the image, not the accompanying textual request.

We identify API-level refusal/safety flags and blocking errors, then parse returned JSON descriptions. It treats empty descriptions, configured sentinel values, and refusal indicators such as reject, refuse, and content filter as refusals. Non-refused descriptions are compared with their clean-image counterparts using the local Qwen3-VL similarity judge.

M2. Editing and Imitation

We further provide quantitative evaluations of image editing and imitation across eight systems. Using the same protected images without task-specific optimization, ADVO achieves 92.1% protection for editing and 91.5% for imitation. Table R5 at the end of this response reports the breakdown of refusal/blocking and failed generation, demonstrating effective protection across both commercial and open-source systems.

M3. Citation Style

We will consistently use parenthetical citations in the revision.


Appendix: Supporting Data and Evaluation Protocol

Table R4. Baseline Comparison Under Official Defaults and Matched Encoders

MethodDefault Avg. PSR ↑Matched-Encoder Avg. PSR ↑
AnyAttack21.519.6
M-Attack33.331.3
FOA-Attack42.557.4
MPCAttack69.179.1
ImageProtector2118.8
ADVO (Ours)89.289.2

Default = official baseline encoders/iterations/other settings, with ϵ=16/255\epsilon=16/255 for all methods. Matched-Encoder = replacing only the baseline proxy encoders with ADVO's seven encoders; original objectives, iteration counts, and other default settings are retained.

Table R5. Quantitative Protection against Image Editing and Imitation

TaskSystem TypeClean Success (%)Refusal / Blocking (%)Failed Generation (%)Total Protection (%)
EditingOpen-source95.70.088.088.0
EditingCommercial96.289.45.294.6
ImitationOpen-source95.70.086.786.7
ImitationCommercial95.888.26.294.4

Results are averaged over three open-source pipelines and five commercial systems, with 100 valid image pairs per system and task. Refusal/blocking and failed generation are mutually exclusive.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.