acceptodds
Under review as a conference paper at ICLR 2027

MIRAGE: Raising the Cost of Malicious Image Editing via False Moderation

Abstract

The proliferation of AI-powered image editing systems raises serious concerns about personal images being arbitrarily manipulated at scale with minimal effort and a lower entry barrier. Prior works on image immunization require access to the model weights and the image editing prompt, which significantly limits their use, especially against powerful commercial black-box image editors such as GPT-Image, Gemini Flash Image (Nano Banana), and Grok Imagine. To address this, we take a system-level view of the problem and identify a previously unexplored protection surface common to most major commercial systems: pre-generation safety moderation. Rather than directly targeting the image-editing model, we target these moderation classifiers to flag immunized images as policy-violating, triggering an automatic refusal. We operationalize this by adding adversarial perturbations to align our image to policy-violating concepts in the representation space. Under the same level of perturbation, our approach, MIRAGE, significantly improves immunization success rate from less than 15% achieved by the best prior technique to more than 88% on multiple closed-source image editing APIs. Our approach is simple, prompt-agnostic, and effective, offering a practical path towards protecting personal images from unauthorized AI-powered editing.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
image immunization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_oZfgOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This study introduces MIRAGE, a model agnostic defense strategy to safeguard personal visual content against unatuhorized editing agents, more specifically closed-source commerical APIs via 'false moderation'. Cleverly, rather than attacking the large model itself, they optimizes the perturbations using an ensemble of open-source encoders and global-local views to align benign images with policy-violating concepts, triggering pre-generation safety filters to refuse the edit. Results across 4 commercial APIs on benchmarks show largely increased success rates compared to baselines

Critical strengths
  • I think the paper is valuable in the sense that it focuses on the highly challenging problem of safeguarding against closed-source models (APIs, Gemini, etc.).

  • I am giving a weak acceptance, as there is comprehensive evaluation showing that the method works on commercial closed-source models, and the topic is novel.

  • It is model- and prompt-agnostic, which makes it an even more powerful approach, and it can be scaled to multiple resolutions.

Critical weaknesses
  • The motivation comes from the fact that these models have safety policy checkers and that their representation spaces are close to open-source ones—so where does this intuition come from? I think an analysis of "why" and "how" this transferability happens is lacking.

  • However, I think the blocker for a clear acceptance is the lack of a comprehensive analysis of the transferability between open-source classifiers and safety checkers in LLMs.

  • Image degradation becomes more visible depending on the budget (see 16/255).

  • It relies on only a single policy category, which is sexually explicit content, whereas there are other ones to explore.

Questions for authors
  • How did you choose the evaluation set of 48 images?
4 comments

We thank the reviewer for the insightful feedback and address each comment below.

An analysis of "why" and "how" this transferability happens

We believe that the success of MIRAGE is due to using an ensemble of diverse image embedding models in our optimization procedure. We provide intuitive and empirical backing for this next.

(Why an ensemble helps) We believe that one of the main contributors to MIRAGE’s empirical success is that we use a large ensemble of open-source models for computing the adversarial perturbations. Past work in adversarial attacks [1][2] has shown that adversarial transfer improves as the perturbation is optimized against more and diverse surrogates. We also see similar results in Fig 5(b) of our manuscript, where immunization rate improves with surrogate size. We conjecture that optimizing against an ensemble leads to more “natural” adversarial examples which can transfer to other classification models.

(Why ensemble of image embedding helps) Prior work in transferability of adversarial examples, e.g., [3], showed that an adversarial example transfers better if (i) the surrogate model is simpler (or smoother); and (ii) the gradients are better aligned between the surrogate and the target models. For (i), our intuition for MIRAGE was that our loss of choice with an ensemble of embedding models is smoother than a loss that involves a white-box image editor or a single white-box moderation model. For (ii), we would attribute a part of this transferability of CLIP-like image embedding to the vast internet-scale data that CLIP models are trained on. It is more likely that the concept of “sexually explicitness” is a fundamental property of the internet-scale data, that is shared across models. We also provide some empirical evidence around this next.

Emprical Evidence - Based on the reviewer's suggestion, we also conduct experiments with the baseline of transferring adversarial examples between white-box and closed-source moderation models. To study this, we run an experiment where we only optimize the adversarial perturbation on ShieldGemma2, an open-source moderation model. We report the immunization rates on SHHQ below

Target modeleps=8eps=16
ShieldGemma294%100%
OpenAI16.7%31.3%
Gemini6.3%6.3%
Grok6.3%20.8%

We find that the rate of transfer of these examples from an open to closed source moderation model is much lower than that of the ensemble achieved by MIRAGE. Even Google’s Gemini moderation seems to be quite different from ShieldGemma2.

To further investigate the source of transferability of MIRAGE, we also perform a similar experiment using a single CLIP model (as opposed to an ensemble of embedding models) to optimize against, and the results are the following

Target Modeleps=8eps=16
OpenAI41%55%
Gemini16.7%22%
Grok14.6%27%

These rates are slightly higher than optimizing using ShieldGemma, which is evidence in favor of our hypothesis that attacks using CLIP-like embeddings are more effective, possibly due to the vast training data these models were trained on (point (ii) above). However, the immunization rates obtained are much lower than using ensembles (Fig 5 of our paper) and lends some support to our claim that attacking an ensemble of models leads to better transferability (point (i) above). Further, we note that there aren’t many open-source multi-modal moderation models available, hence we could not conduct an experiment against an ensemble of multiple ShieldGemma like models. These experiments further motivate our choice of using an ensemble of embedding and moderation models for MIRAGE and provide some intuition on what is essential for the attack to transfer to close source models.

Single policy category

In our ablation (Figure 5 (d)) study we choose the target category to be sexually explicit content as well as violent content, and find that the former works better empirically. In our preliminary experiments, we also looked at other categories in OpenAI’s moderation policies including copyrighted content, hate speech and bio-weapons, but we found that these categories were not effective. We will add these findings in the revision.

How did you choose the evaluation set of 48 images?

We balanced the gender and the background in the images (some images contained a white background while others were in a more natural setting). We then randomly selected images while keeping an equal balance across these categories. In particular, the selection of the images was done before running any of our methods on the images. We will make it more explicit in the revision.

Thank you for the extra experiment you have conducted, I understand how an ensemble helps in transferability, my questions are resolved in that part.

I would like to see some results on other contents if you can provide any (not only sexually explicit content), because it will support the model's generalization capabilities, if you can provide more results on other categories and update the paper, I would reconsider my score.

Thanks

AuthorsOpenReview ↗

Official Response

We thank the reviewer for their response.

In the table below, we show immunization rates at ϵ=16255\epsilon=\frac{16}{255} for four different target categories --- sexually explicit imagery, graphic violence, hate symbols (e.g. a swastika) and copyrighted content (e.g. the likeness of Mario). These are against the usage policies of OpenAI, XAI and Google. The immunization rates using such images as targets are ---

APISexually ExplicitGraphic ViolenceHate SymbolsCopyrighted
OpenAI100%73%98%98%
Gemini100%27%83%0%
Grok100%96%0%29%

As we see, different providers exhibit different levels of moderation sensitivities across concepts and categories. However our perturbations do trigger refusal across multiple API providers, indicating that they generalize to some extent.

We have also updated our manuscript, adding the extra experiments and analysis from the rebuttal period. The changes are marked in blue.

Let us know if there are any lingering questions that we can answer for you to reconsider your score, we would be happy to provide any further clarifications!

AuthorsOpenReview ↗

Gentle Reminder

As the response period draws to a close, we are gently following up on our response, and would be happy to provide any clarifications for you to reconsider your evaluation of our work! Thanks again for your insightful review and engagement!

Reviewer_8FkZOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a black-box immunization method against malicious image editing. Most prior work either optimizes directly against white-box models or applies black-box optimization against proprietary VLMs, both of which yield poor results against commercial API models such as Google’s Nano Banana, OpenAI’s GPT-Image, and Grok Imagine. The paper instead targets the content moderation classifier deployed by all such platforms. This classifier evaluates both the input and the output images and withholds the response if either image triggers a safety concern (e.g., explicit or violent content).

The optimization relies on an ensemble of white-box surrogate embedding models and optimizes the image at two resolutions: a global view, in which the entire image is resized to match the input dimensions expected by the ensemble, and a local view, in which patches of the same resolution are used to avoid performance degradation when mapping back to the original resolution. The approach works well as it essentially crafts a transferable adversarial example, and prior work has shown such examples to transfer well when optimized against an ensemble of surrogate models.

Critical strengths
  • Targeting the content moderation classifier rather than the image editing model itself is a highly interesting idea. The approach is simple and intuitive, yet effective and fairly robust, since such content moderation filters can safely be assumed to be deployed in all standard commercial image editing tools.

  • The paper is well written, clear, and well organized. I enjoyed reading it.

  • I especially appreciated the social media evaluation, as this is a concern that immediately comes to mind. The results initially surprised me, since I expected some of the images to be rejected. But in a way this should have been expected, as social media platforms are careful to avoid false positives, which can drive users away from the platform, whereas image editing tools can tolerate more false positives and focus instead on false negatives, which pose a greater risk to the provider (e.g., legal consequences or reputational damage).

  • The approach appears highly effective. Combining global and local views is a simple and effective way to handle the resolution mismatch.

Critical weaknesses
  • The paper evaluates only a single prompt. While the results are already convincing, it would be valuable for completeness to show that the method does not overfit to this specific prompt and works equally well with others.

  • The main limitation is that the immunization can be removed relatively easily. The paper evaluates a broad set of adaptive attackers, and the results show that, with some computational effort, an attacker can effectively remove the protection. While not every adversary will invest this effort, a determined adversary can do so with reasonable resources. Moreover, because the attacker receives a refusal when editing an image that is clearly benign, they can immediately infer that the image is protected and decide whether to invest in removing the protection, which the results in the appendix show to be feasible.

  • Having an image flagged as unsafe by content moderation filters can have consequences. While the authors convincingly show that this does not affect social media platforms, other AI-based tools that process the image may block it as well. For example, if the image owner publishes an article or blog post containing an immunized image, readers who ask their AI assistant to summarize or even find the article may be refused.

  • As discussed above, content moderation filters in image editing platforms appear to be much more sensitive to false negatives than to false positives, so it is reasonable to assume they are not trained to be robust against adversarial examples. However, some systems may deploy more robust classifiers. I would have liked to see an evaluation against a system with a more robust classifier. This would necessarily be a simulated setting that mimics a deployed system, since the content moderation of real systems cannot be replaced.

Questions for authors

Could you evaluate your approach against content moderation classifiers that were explicitly trained to be robust against adversarial examples?

2 comments
AuthorsOpenReview ↗

Response to Review

We thank the reviewer for their insightful comments. We address some of their concerns below

Could you evaluate your approach against content moderation classifiers that were explicitly trained to be robust against adversarial examples?

Thank you for this insightful suggestion. To evaluate our method against robust content classifiers, we turn to ShieldGemma2. According to page 4 of the ShieldGemma2 tech report, the model was trained on data which was adversarially generated to produce false positives and false negatives for moderation. Since this model was a part of our original 8-model ensemble, we re-ran MIRAGE with a 7-model ensemble without ShieldGemma2. We did not see a drop in the immunization rates against OpenAI, Grok or Gemini models for these resulting immunized images at $\epsilon=8/255 $ and ϵ=16/255\epsilon=16/255. We then checked the refusal rates on ShieldGemma2. Since this model returns a probability that an image is unsafe, we used a threshold of 0.5 to call an image flagged (we will also show the histogram of scores in our updated manuscript). The refusal results are below

Modeleps=8 refusalseps=16 refusals
OpenAI89.6%100%
Gemini75%100%
Grok77.1%100%
ShieldGemma291.6%97.9%

We find that the immunization rate against ShieldGemma2 is similar to other moderators, suggesting that MIRAGE works even against such robust moderation models.

The paper evaluates only a single prompt

We want to note that our 100 image dataset has unique prompts for each image. We now also have evaluated immunization against 4 other unique prompts for each SHHQ image and find no difference in immunization rates. We will report this new finding in the revision. Finally, pushing our robustness against prompts to an extreme, in Figure 4, we showed that MIRAGE achieves 100% immunization rate against an adversary which does prompt optimization, which by construction tests our method against a diverse set of prompts that are maliciously optimized per image.

Adaptive adversary

We agree with the reviewer that a motivated adversary can bypass our immunization scheme by investing some compute. For example, the CLIP attack in App D.2.2 was run on a single A40 GPU, taking ~4 minutes per image. Note that in order to run this attack, an adversary requires access to powerful hardware, which we argue raises the cost of malicious editing.

Further, an attack like this is not completely free, since it leads to degraded image quality. To measure this, we looked at the LPIPS distance between the original image and the image after the CLIP attack from Table 5. For the immunized images, the LPIPS distance from the original image was 0.3394 on average. This increases to 0.4779 on average for the attacked image, indicating a degradation in perceptual quality. But perhaps more importantly, this also affects downstream edits, e.g. the identity of the person in the image changes in 47% of the images (judged by a VLM) after edits by the OpenAI model. This change is lower for Gemini and Grok, at 14% and 12.5% resp

other AI-based tools that process the image may block it as well.

This is an interesting point, and we agree that our method might have such downstream effects. To study this, we also did a preliminary experiment, where we used the same images from SHHQ in an image captioning task. We present the following results.

TaskModelRefusalsRefusal Type
CaptioningGPT-4.1100%Soft
CaptioningGPT-6-Luna0%NA
EditingGPT-Image-1.5100%Hard
EditingGPT-Image-2.6100%Hard

Here Soft refusal refers to the model outputting a text along the lines of “Sorry, I cannot help with that request”, while a Hard refusal refers to the API returning an error. From the results above, it seems that the moderation system for image-to-image models is different from the moderation for image-to-text models. In fact, we find that even on our unsafe target images, GPT-6-Luna produces zero refusals on the captioning tasks. We conjecture that image-to-text models do not have a separate moderation pipeline, but it is a part of the model itself, which leads to fewer refusals with more advanced models. This does not appear to be true for image-editing models as the immunization rates seem to be similar in our preliminary experiments. We suspect that this is related to how malicious image editing can be more harmful compared to malicious image captioning (or summarization), and the model provider’s desire to be more compliant to policies around image editing. However, we agree that it would be interesting future work to analyse and characterize these differences. We will include these findings in the revision.

Thank you, this addresses most of my concerns and I will keep my positive score.

Reviewer_dR5aOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MIRAGE, an image immunization method that protects personal images against unauthorized editing by exploiting the content moderation mechanisms of commercial image-editing systems. MIRAGE optimizes adversarial perturbations to align the protected images with policy-violating concepts in the representation spaces of open-source surrogate models, thereby inducing content moderation refusals.

Critical strengths
  • A novel and practically relevant system-level perspective on image immunization. Unlike prior methods that directly target image generation models, MIRAGE exploits the shared content moderation mechanisms of commercial image-editing systems.
Critical weaknesses
  • Limited methodological novelty and insufficient differentiation from closely related work. The core idea and technical components of MIRAGE appear closely related to existing research. In particular, [1] has already explored protecting personal images by inducing refusal behaviors in multimodal models, [2] demonstrates that CLIP-based surrogate attacks can manipulate the safety-related behavior of proprietary multimodal systems. Moreover, MIRAGE largely builds upon existing techniques, including surrogate model ensembles, feature alignment, PGD, and global-local optimization[3].
  • Insufficient evidence for the practical security benefits against adaptive adversaries. The central motivation of this paper is to raise the cost of malicious image editing. However, MIRAGE appears vulnerable to relatively straightforward adaptive attacks, as shown in Table 5.
  • Poor paper organization and insufficiently focused presentation. The paper would benefit from substantial improvements in its organization and presentation. In particular, the motivation is repeatedly discussed in the Introduction and Section 4.1, while the distinction between the central methodological contribution and auxiliary implementation techniques remains unclear.

[1] Shao, Zedian, et al. "Leave My Images Alone: Preventing Multi-Modal Large Language Models from Analyzing Images via Visual Prompt Injection." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

[2] Zhang, Jie, et al. "Laundering AI Authority with Adversarial Examples." arXiv preprint arXiv:2605.04261 (2026).

[3] Zhao, Xiaohan, et al. "M-Attack-V2: Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting." (2025).

Questions for authors
  • Novelty over closely related work. Could the authors clarify the conceptual and technical contributions of MIRAGE beyond existing refusal-inducing image protection methods?
  • Practical effectiveness against adaptive attacks. Could the authors report the computational cost and end-to-end editing success rates of the adaptive attacks in Appendix D.2.2, including the perceptual quality and identity preservation of the resulting images? This would help quantify whether MIRAGE meaningfully raises the cost of malicious image editing.
  • Contribution and efficiency of gradient caching. Could the authors provide an ablation study of model dropout and secant gradient caching, including comparisons with standard gradient accumulation or simpler caching strategies in terms of runtime, GPU memory, and immunization effectiveness?
2 comments
AuthorsOpenReview ↗

Response to review (1/2)

We thank the reviewer for their comments and suggestions, and are glad they found our work to be novel and practically relevant. We address some of their concerns below.

Limited methodological novelty and insufficient differentiation from closely related work

We thank the reviewer for the pointers to the closely related works. We would like to clarify that the setting and motivation of [1] and [2] are different from our work.

Both MIRAGE and [1] share the high level goal of triggering refusal, but their similarity stops there. In particular, the setting and the methodological techniques are different. Because [1] is focused on immunizing against visual question answering by local VLM under white-box access, the proposed method directly optimizes to trigger refusal tokens using back-propagation. This optimization is specific to the model being protected against. This results in the perturbation that (i) protects against specific queries and (ii) does not transfer to other models. On the other hand, we focus on AI-powered image editing systems with API access, where we cannot apply optimization methods directly. This requires a novel approach of using concepts from the terms-of-use to design perturbations to trigger the moderation, taking a system-level view of the problem. Such designed perturbation has the added benefit of being prompt agnostic and universally working for several proprietary systems.

[2] attacks commercial, black-box VLMs. However, the attack objective is closer to standard jailbreaking tasks, where the adversary aims to construct perturbations which can bypass the moderation and decision making pathways of commercial VLMs. This enables the production of output that is policy-violating, e.g. producing fake news about real people. Our work on the other hand perturbs input images which induce content moderation filters to refuse outputs on otherwise “benign” images. Our method also has technical differences from this work, and using the exact technique from [2] has lower immunization rates in our setting as we show in Fig 5. We discuss this difference in our extended related work in Appendix A.

Finally, our method build on advances in transferable adversarial examples and hence has similarities to [3], with some differences –

  1. In addition to the multi-crop optimization from [3] we also add a global view of the image (Eq (2) of our paper) while computing the adversarial perturbation.
  2. We introduce model dropout and gradient caching (eq (3) of our paper) which increases the efficiency of our method by 2.86x in terms of runtime with no drop in immunization rates.

We would like to emphasise that the main contribution of our work is to take a system-level look at image immunization and construct transferable perturbations to falsely trigger content moderation and protect against black-box image editing.

Insufficient evidence for the practical security benefits against adaptive adversaries, including the perceptual quality and identity preservation of the resulting images

We acknowledge the reviewer’s constructive feedback. For the computational costs, the CLIP attack was run on a single A40 GPU, taking ~4 minutes per image. Note that in order to run this attack, an adversary requires access to powerful hardware, which we argue raises the cost of malicious editing.

To measure perceptual quality, we looked at the LPIPS distance between the original image and the image after the CLIP attack from Table 5. For the immunized images, the LPIPS distance from the original image was 0.3394 on average. This increases to 0.4779 on average for the attacked image, indicating a degradation in perceptual quality. But perhaps more importantly, this also affects downstream edits, e.g. the identity of the person in the image changes in 47% of the images (judged by a VLM) after edits by the OpenAI model. This change is lower for Gemini and Grok, at 14% and 12.5% resp.

Response continues below

AuthorsOpenReview ↗

Response to Review (2/2)

Ablation study of model dropout and secant gradient caching

In line with the reviewer’s suggestion, we ran an ablation study for this component of our method. We find that using model dropout is empirically 2.86x faster in terms of runtime as compared to using all models at every optimization step. The GPU memory consumption is similar for both the alternatives, and the immunization rate is also the same. This indicates that model dropout leads to greater computational efficiency at essentially no performance cost. We will revise the manuscript to include these new results.

We also performed an ablation study with secant caching. As an alternative, we considered simply using the last gradient (git−1g_i^{t-1} from eq (3)). This empirically has similar runtime and memory consumption as secant caching. We find that this leads to a slightly lower immunization rate at \eps=8/255\eps=8/255 as shown below

MethodGrokOpenAIGemini
Last Gradient75%89.6%70.8%
Secant Caching77.1%89.6%75%

We will include these ablation studies in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.