acceptodds
Under review as a conference paper at ICLR 2027

IRIS: Imperceptible, Robust Image Sealing for Cooperative Anti-Edit Protection

Abstract

Unified image-editing models such as FLUX.2 unlock substantial creative potential, but also make publicly shared content easy to repurpose without consent. For example, they can fabricate images of a person in scenarios that never occurred or imitate an artist's signature style. Existing input-only defenses add subtle perturbations to disrupt such edits, but this small signal must survive everyday image processing. Simple operations such as cropping, flipping, resizing, or compression can therefore weaken protection, while increasing perturbation strength can compromise visual quality. We introduce IRIS, a cooperative defense in which the model provider participates in enforcing the image owner's opt-out. This cooperation allows IRIS to separate robust signal design from protective response learning. It embeds subtle, image-dependent signals using a frozen watermarking model and fine-tunes the editor with a target-image-free objective that selectively cools down the velocity field on protected inputs, keeping generation in the noisy regime while preserving clean editing. Creators can apply the protection to new images without per-image optimization or access to the editor, and the same protected image can be used across cooperating models. Experiments with FLUX.2 and Qwen-Image-Edit demonstrate this selectivity: IRIS combines low perceptual distortion and strong protection under everyday image processing with largely preserved clean-image editing fidelity.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
protection, anti-editing, defense, image cloaking, adversarial

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Rq2gOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a cooperative framework that protects images from unauthorized editing through imperceptible watermarks and a learned protective response.

It fine-tunes the editor to suppress its generative velocity on watermarked inputs while preserving clean-input behavior through teacher matching, without requiring paired editing data or predefined output targets. Experiments demonstrate strong protection under common image transformations with limited degradation of normal editing quality. However, protection depends on using an adapted editor and remains vulnerable to diffusion-based purification at native watermark strength.

Critical strengths
  • The paper is well written and easy to understand. The proposed training objective is simple and effective, the training is not costly and works well.
Critical weaknesses
  • The biggest concern is that this paper seems to have very tiny contribution. The only difference between this paper and other backdoor paper is a zero regression loss to surpress watermarked input and clean teacher distillation.
Questions for authors

Can you show more about how this paper is different from the literature?

3 comments
AuthorsOpenReview ↗

Relation to Prior Work and Purification (1/2)

We thank the reviewer for highlighting the clarity of our paper and the low training cost of IRIS. We agree that the objective is simple. Our intuition was that it would be easier to learn a selective response if the editor had to predict the same velocity on every protected input. We appreciate the opportunity to elaborate more on the intuition and reasoning behind this design choice.


Q1/W1: How does IRIS differ substantively from the existing backdoor literature?

GuardDoor and Chen et al. established the cooperative protection setting. We built on that but when we applied their approaches to FLUX.2, both failed to preserve editing on unprotected images. Each completed only 44.3% of requests on unprotected photos, compared with 88.3% for IRIS (Table R5; revised paper, Table 1). Our controlled ablations in Table R6 show that the choice of the response objective, i.e. what we teach the editing model to do for protected images, has an independent effect on selectivity when the other training choices are held fixed. Our methodological contribution is to show that the safeguard response we teach the editor on protected photos matters for how well it can preserve ordinary editing abilities. GuardDoor trains the encoder toward the features of a fixed target, whereas Chen et al. train the editor to generate a predefined image (Appendix B.2).

Table R5. VLM edit success of IRIS in comparison to full cooperative baselines on 300 images.

EditorMethodClean ↑Protected ↓Selectivity ↑
FLUX.2Original90.3%90.3%–
GuardDoor44.3%18.0%26.3%
Chen et al.44.3%0.0%44.3%
IRIS (Ours)88.3%0.3%88.0%
Qwen-Image-EditOriginal84.0%84.0%–
GuardDoor35.7%34.3%1.4%
IRIS (Ours)86.0%0.0%86.0%

The earlier work already showed that a trigger can change a model’s behavior. Our question was which response to teach a rectified-flow editing model so that it can still follow normal unprotected requests. To answer it, we kept the watermark and clean training loss fixed and changed only the response on protected images (Table R6 below).

The flow equation helps explain why the choice of response matters. Along the training path xt=(1−t)y+tϵx_t=(1-t)y+t\epsilon toward a replacement image yy, the target velocity at t>0t > 0 is v∗(xt,t)=ϵ−y=(xt−y)/tv^{*}(x_t,t)=\epsilon-y=(x_t-y)/t. Even for the same fixed image, the editor must predict a different velocity as the noise sample changes. IRIS asks for v∗=0v^{*}=0 on every protected input. Since each sampling update scales with the predicted velocity, a zero prediction stalls denoising without requiring the editor to synthesize a particular replacement. Our intuition is that this simpler velocity field cooldown objective makes it easier to learn when to suppress an edit. Table R6 shows a controlled comparison of different response objectives measuring the resulting selectivity across both editors.

Table R6. Response objectives on FLUX.2 and Qwen-Image-Edit, averaged over four transformations. Cells show VLM edit success (%) with DINO similarity (×100) in parentheses. Selectivity is clean minus protected edit success in percentage points (revised paper, Table 3).

Training objectiveFLUX.2 clean ↑FLUX.2 protected ↓FLUX.2 selectivity ↑Qwen clean ↑Qwen protected ↓Qwen selectivity ↑
Zero velocity (IRIS)88.0% (93.6)1.25% (3.0)86.75%86.7% (84.7)0.0% (−0.4)86.7%
Gray target (GuardDoor-style)82.4% (78.5)78.8% (77.5)3.6%83.7% (92.7)83.4% (91.4)0.3%
Return original source90.7% (89.7)76.4% (81.1)14.3%81.1% (89.1)80.0% (88.2)1.1%
Fixed noise target82.9% (78.5)54.2% (52.6)28.7%82.8% (90.7)83.0% (89.7)−0.2%
Fixed target + image loss (Chen et al. style)7.7% (11.4)5.7% (9.8)2.0%83.4% (89.9)82.1% (88.7)1.3%
AuthorsOpenReview ↗

Relation to Prior Work and Purification (2/2)

“Protection [...] remains vulnerable to diffusion-based purification at native watermark strength”

Thank you for pointing out that diffusion purification weakens IRIS at its native watermark strength. We agree, and the revised paper now makes clearer how much protection remains and how that changes when the watermark is strengthened. After DiffPure (t=50t=50), protected images remained harder to edit with native-strength IRIS: edit success was 4.3% lower on protected photos than on clean photos for FLUX.2 and 10.6% lower for Qwen. GuardDoor lost nearly all selectivity under the same conditions. On both editors, it fulfilled only about a third of requests on either clean or protected photos (Appendix D.2, Tables 17 and 18).

However, IRIS protection can be restored cheaply at inference time: doubling the perturbation strength, without retraining, blocks 94% of edits on FLUX.2 and 100% on Qwen after purification at t=50t=50 (Table 11). Input LPIPS rises from 0.042 to 0.113, which is still below the values reported for EditShield and PhotoGuard. In the FLUX.2 example in Figure 7, a stronger watermark kept the requested edit from succeeding even after stronger DiffPure at t=150t=150, and concentrating the added signal in high-frequency, textured regions reduced input distortion relative to the stronger uniform setting. The revised main results report the t=50t=50 amplification comparison on 100 VetoBench-General photos.


Thank you for prompting us to make this distinction clearer. The revised PDF explains this intuition in Section 3.2 and reports the controlled comparison in Appendix Table 3.

Dear Reviewer, as the discussion period is coming to an end, we would appreciate hearing from you. Thank you!

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

IRIS uses an existing watermarking model to embed trigger signals and fine-tunes an image editor to selectively suppress generation on protected inputs. It trains the editor’s velocity predictions toward zero for watermarked images while preserving normal editing behavior on clean inputs. Experiments demonstrate the effectiveness of this cooperative anti-editing approach under the evaluated conditions.

Critical strengths
  1. The paper addresses a meaningful problem: selectively preventing edits to protected images while preserving the model’s editing capabilities on clean inputs. Maintaining this distinction is important for the practical usefulness of the defense.
  2. The method is simple, using a zero-velocity objective to suppress generation on protected inputs. Extensive experiments and ablation studies provide empirical support for its effectiveness.
Critical weaknesses
  1. Limited conceptual novelty. The paper builds on existing cooperative protection and watermark-triggered response frameworks, using a pretrained watermark embedder to introduce imperceptible signals and fine-tunes the editing model to disrupt generation for protected inputs while preserving normal behavior on clean inputs. Similar ideas have already been explored in GuardDoor [1] and Invisible backdoor triggers [2]. The paper does not introduce a new watermarking method. Its main modifications are to regress the velocity predictions for protected inputs toward zero and preserve clean-input behavior through teacher matching, using a training procedure that does not require paired editing data. In my view, these changes represent a limited methodological extension, and their conceptual significance is not sufficiently established.
  2. Insufficient evidence for the independent contribution of the proposed loss. The current experiments do not clearly isolate the contribution of the new loss function. The observed performance may jointly reflect the robustness of the pretrained watermark, adaptation of the generative backbone rather than the encoder, clean-input teacher supervision, and the zero-velocity objective. Further controlled ablations are needed to disentangle these factors and establish whether the proposed loss design provides a distinct and substantive benefit.

[1] Guarddoor: Safeguarding against malicious diffusion editing via protective backdoors, 2025

[2] Invisible backdoor triggers in image editing model via deep watermarking, 2025

Questions for authors
  1. Could the authors clarify the key methodological insight of IRIS beyond the existing cooperative protection and watermark-triggered response frameworks, particularly GuardDoor [1] and Invisible backdoor triggers [2]? Specifically, what explains the advantage of zero-velocity regression over fixed-target responses, and what new understanding does this design provide about achieving selective protection while preserving clean editing capabilities?
  2. Table 3 provides useful controlled comparisons, but reports only DINO similarity, which does not directly measure editing success. Could the authors evaluate using task-level editing-success judgments, to establish whether the advantage of zero-velocity regression persists at comparable clean-editing utility?
5 comments
AuthorsOpenReview ↗

Key Insight and Component Ablations (1/3)

Thank you for highlighting how important it is to preserve ordinary editing and for asking us to isolate the contribution of the response objective. We appreciate this opportunity to clarify the key methodological insight over the prior work and how it connects to our zero-velocity objective of IRIS.


Q1a / W1: What is the key methodological insight beyond GuardDoor/Chen?

GuardDoor and Chen et al. established the cooperative protection setting. We built on that, but when we applied their approaches to modern FLUX.2, both failed to preserve editing on unprotected images. Each completed only 44.3% of requests on unprotected photos, compared with 88.3% for IRIS (Table R2; revised paper, Table 1). Our controlled ablations in Table R3 show that the choice of the response objective, i.e., what we teach the editing model to do for protected images, has an independent effect on selectivity when the other training choices are held fixed.

For an extended comparison, we adapted Chen et al.’s full supervised training procedure to FLUX.2 on 10,000 image pairs and show the results in Table R2. We also tested their fixed-target response objective separately in a controlled ablation in Table R3. The full method blocked 100% of protected edits but completed only 44.3% of unprotected edits. On the same photos, IRIS blocked 99.7% of protected edits and completed 88.3% of the unprotected ones, which is close to the original model’s edit success rate of 90.3% while achieving an almost perfect protection rate.

Our methodological contribution is to show that the safeguard response we teach the editor on protected photos matters for how well it can preserve ordinary editing abilities. With the watermark and clean-preservation training held fixed, our proposed zero-velocity approach gave the best balance between blocking protected edits and preserving ordinary editing on both FLUX.2 and Qwen-Image-Edit (Table R3). Below, we explain how a constant velocity target differs from the target required to generate a fixed image, and why this distinction may help the editor respond selectively.

Table R2. VLM edit success of IRIS in comparison to full cooperative baselines on 300 images.

EditorMethodClean ↑Protected ↓Selectivity ↑
FLUX.2Original90.3%90.3%–
GuardDoor44.3%18.0%26.3%
Chen et al. (full FLUX.2 adaptation)44.3%0.0%44.3%
IRIS (Ours)88.3%0.3%88.0%
Qwen-Image-EditOriginal84.0%84.0%–
GuardDoor35.7%34.3%1.4%
IRIS (Ours)86.0%0.0%86.0%

Q1b: Why does zero-velocity regression help relative to fixed-target responses?

The key difference is what the editor has to predict on a protected input. Suppose we want it to produce a fixed replacement image with latent yy. Under the rectified-flow parameterization in Section 3.2, the training path is xt=(1−t)y+tϵx_t=(1-t)y+t\epsilon, where ϵ\epsilon is the sampled noise. The target velocity is therefore v⋆(xt,t)=ϵ−y=(xt−y)/tv^{\star}(x_t,t)=\epsilon-y=(x_t-y)/t for t>0t > 0. The replacement image is fixed, but the velocity needed to reach it depends on the current latent and timestep. At inference, the seed sets the initial latent, and the prompt and reference image guide the trajectory from there. Even with the same seed, changing the instruction or reference leads to a different latent at the same timestep, so reaching the same replacement image always requires a different velocity. The editor must learn this complex replacement behavior on protected photos while continuing to follow ordinary editing instructions on clean ones.

With zero velocity, the target is the same at every protected state, regardless of the seed, prompt, reference image, or point along the trajectory. Predicting zero cools down the velocity field and stalls generation. Our intuition is that this simpler response lets the editor devote more of its capacity to learning when to suppress an edit, which may help it preserve normal editing of unprotected images.

Table R3 measures the effect directly. On FLUX.2, the fixed-noise target lowered clean edit success by 5.1 percentage points relative to zero velocity. With the gray target, the gap between clean and protected edit success narrowed from 86.75 to 3.6 percentage points (Appendix A.2, Table 3).

AuthorsOpenReview ↗

Key Insight and Component Ablations (2/3)

Q2: Does Table 3’s advantage persist under task-level editing-success judgments at comparable clean-editing utility?

Yes. We incorporated your feedback and extended all tables in the paper to include the VLM judgment metric. As reported in the revised main results (Table 2), on Qwen-Image-Edit the editor completed 86.7% of clean requests with zero velocity and 83.7% with the gray target. On protected photos, it completed 0.0% and 83.4%, respectively (Table R3). We saw the same pattern on FLUX.2. The revised paper reports task-level edit success alongside DINO similarity throughout the editing tables. We also included training the model to simply return the original photo as an additional baseline.

Table R3. Response objectives on FLUX.2 and Qwen-Image-Edit. VLM edit success (%) is averaged over four transformations, with DINO similarity (×100) in parentheses. Full results appear in Table 3 of the revised paper.

Training objectiveFLUX.2 clean ↑FLUX.2 protected ↓FLUX.2 selectivity ↑Qwen clean ↑Qwen protected ↓Qwen selectivity ↑
Zero velocity (IRIS)88.0% (93.6)1.25% (3.0)86.75%86.7% (84.7)0.0% (−0.4)86.7%
Gray target (GuardDoor-style)82.4% (78.5)78.8% (77.5)3.6%83.7% (92.7)83.4% (91.4)0.3%
Return original source90.7% (89.7)76.4% (81.1)14.3%81.1% (89.1)80.0% (88.2)1.1%
Fixed noise target82.9% (78.5)54.2% (52.6)28.7%82.8% (90.7)83.0% (89.7)−0.2%
Gray target + image loss (Chen et al. style)7.7% (11.4)5.7% (9.8)2.0%83.4% (89.9)82.1% (88.7)1.3%
AuthorsOpenReview ↗

Key Insight and Component Ablations (3/3)

W2: How much of the gain comes from the watermark, the adapted module, clean preservation, and the response objective?

Each of the four components mattered. In the fixed-noise and zero-velocity runs in Table R3, we kept the watermark and the clean-preservation weight fixed. Both used the same loss composition: L=12,MSE(vθclean,v0clean)+12,MSE(vθprotected,v∗)\mathcal{L}=\tfrac{1}{2}\\,\mathrm{MSE}(v_\theta^{\mathrm{clean}}, v_0^{\mathrm{clean}})+\tfrac{1}{2}\\,\mathrm{MSE}(v_\theta^{\mathrm{protected}}, v^{*}).

Here vθv_\theta is the adapted editor’s velocity prediction and v0v_0 is the frozen original editor’s prediction on the same clean input. IRIS sets the target v∗=0v^* = 0, while the fixed-noise baseline sets it to v∗=ϵ−zfixedv^*=\epsilon-z_{\mathrm{fixed}}, where ϵ\epsilon is the sampled noise and zfixedz_{\mathrm{fixed}} is the latent of the fixed noise image. Changing only the protected response target from fixed noise to zero velocity kept ordinary editing strong on Qwen (82.8% to 86.7%) while reducing successful edits on protected images from 83.0% to 0.0%.

The flow-level objective gives us a training target directly on the denoising transformer by specifying zero velocity on protected images. When we adapted only the FLUX.2 image encoder, 76.7% of protected edits succeeded (Table R4a). On Qwen, adapting only the image encoder led to a different failure: none of the 1,200 ordinary edits or 1,200 protected edits succeeded across the four transforms (Table R4b). Adapting the transformer preserved 86.7% ordinary edit success while keeping protected edit success at 0.0%. One possible explanation is that the encoder shapes the image representation only at the input, whereas the transformer uses that representation throughout editing and can learn when to suppress an edit.

The watermark and clean-preservation loss each play a different role. Changing only the signal from WAM to GuardDoor’s reconstruction signal made FLUX.2 less reliable on both kinds of images. Ordinary edit success fell from 88.0% to 22.6%, and successful edits on protected photos rose from 1.25% to 15.9%. Removing clean preservation altogether drove ordinary edit success to 0% on both editors, so, as expected, clean preservation is important to keep the editor useful on clean images. Table R3 shows why the choice of response part of the objective still matters even when the preservation term is kept the same.

Table R4a. IRIS component comparison on FLUX.2 with 300 VetoBench sources across four transforms. Clean and protected columns show VLM edit success in percent, with DINO similarity (×100) in parentheses. Selectivity is clean minus protected edit success in percentage points (higher is better).

ComponentVariantClean ↑Protected ↓Selectivity ↑
Full IRIS (Ours)–88.0% (93.6)1.25% (3.0)86.75%
WatermarkGuardDoor signal in place of WAM22.6% (24.3)15.9% (18.8)6.7%
Adapted moduleImage encoder in place of denoising transformer90.2% (98.9)76.7% (89.6)13.5%
Clean preservationRemove the clean teacher loss0.0% (1.7)0.0% (1.7)0.0%
Response objectiveFixed gray target in place of zero velocity82.4% (78.5)78.8% (77.5)3.6%
Response objectiveReturn the original photo in place of zero velocity90.7% (89.7)76.4% (81.1)14.3%

Table R4b. Qwen-Image-Edit on the same 300 sources and four transforms as R4a. Clean and protected columns show VLM edit success (%) with DINO similarity (×100) in parentheses.

ComponentVariantClean ↑Protected ↓Selectivity ↑
Full IRIS (Ours)–86.7% (84.7)0.0% (−0.4)86.7%
WatermarkGuardDoor signal in place of WAM54.1% (50.0)17.2% (15.5)36.9%
Adapted moduleImage encoder in place of denoising transformer0.0% (−1.5)0.0% (−1.5)0.0%
Clean preservationRemove clean preservation0.0% (−0.2)0.0% (−0.2)0.0%
Response objectiveFixed gray target in place of zero velocity83.7% (92.7)83.4% (91.4)0.3%
Response objectiveReturn the original photo in place of zero velocity81.1% (89.1)80.0% (88.2)1.1%

Thank you for prompting us to clarify the connections to prior work and the conceptual role of the zero-velocity objective for a successful safeguard. We revised the main paper PDF to make the intuition and reasoning behind IRIS more explicit and included the component comparisons from Tables R4a and R4b in Appendix A.5 of the revised paper (Tables 5 and 6).

Reviewer_emB5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes a method (IRIS) to protect images from getting edited with generateive models. The threat model that assumes cooperation between the the model provider and the image owner. The image owner watermarks an image with an off-the-shelf watermarking model. The model-provider finetunes the model to make it predict suppressed velocities for watermarked images and original velocities otherwise. By doing so, the authors show that immunized images provide strong protection against editing even when immunized images are purified via common transformations. The immunized images also undergo low distortion. Meanwhile clean edits remain high quality. The method also allows for inference-time control of noise-perception vs ppurification-robustness tradeoffs. The method was demonstrated on Flux2 and Qwen-Image-Edit models, along with detailed analyses and ablations.

Critical strengths
  1. The proposed method addresses (at least conceptually) some major limitations of most image-editing-immunization lilterature. Namely, one image immunized against multiple (cooperating) editng models and strong purification robustness.
  2. The training method is simple and does not require paired data.
  3. The analysis, details and ablations are quite extensive. The supplementary material is high quality.
Critical weaknesses
  1. Limited study of purification attacks specific to the proposed threat model: The model providers release the IRIS-fine-tuned version of the editor-model. But they must also release the coresponding watermark-model so that user can watermark their images to protect them. In such a scenario, attackers can try watermark reversal/purification under the known watermarking model. Such attacks have been discussed in the appendix but not experminted with or evaluated.
Questions for authors
  1. How would the original watermark be affected if I optimized the watermarked-image with PGD with the objective of removing the watermark. Since the watermark encoder is publicly released under the threat model, perhaps something like min_{x_p} ||WAM_enc(x_p)|| as a purification attack?
  2. Will the method survive if the released model is finetuned again by the attacker with a reverse objective? L_rev = 0.5MSE(v_clean,v_T) + 0.5(v_wm, v_T), where v_T is target velocities obtained from the IRIS-based-model from a non-watermarked image. v_clean and v_wm and are the velocities predicted by the new-attackers'-student model from clean and watermarked images.
4 comments
AuthorsOpenReview ↗

Adaptive Attacks and Purification

Thank you for the constructive review. We appreciate you highlighting the clarity of the paper and that IRIS conceptually solves major limitations of existing image-protection frameworks. We conducted additional experiments on your proposed adversarial attacks and they helped us to deepen our understanding of IRIS.


Q1: Does PGD against the publicly released watermark system remove the signal and restore editing?

PGD disrupted message decoding, but it did not restore editing. To test this, we followed your suggestion and gave PGD the public WAM model and the exact message we used for fine-tuning the model with IRIS. After the PGD attack with a perturbation budget of ϵ=8/255\epsilon = 8/255, IRIS still rejected 99.7% of the protected edits (Table R1; Appendix D.2, Table 16). WAM’s bit accuracy fell to 1.4% on the protected photos. One possible explanation is that IRIS responds to features left by the WAM embedder even when the original message is no longer reliably decoded. This would be consistent with our response-specificity results in Appendix C.4. This observation may open exciting new directions for future work on more message-specific cooperation or specialized watermarking signals for anti-edit protection frameworks.

Table R1. FLUX.2 successful edits before and after PGD (in % of 300).

MethodPGD budget ε / 255Protected edit success ↓
IRIS (WAM)No PGD0.3%
No IRIS489.3%
IRIS (WAM)40.3%
No IRIS887.3%
IRIS (WAM)80.3%

Q2: Does the specified reverse-distillation fine-tuning remove protection from the released editor?

In this scenario, the attacker has full access to the WAM embedder and the exact message to create a protected copy of each of 1,000 training photos. After 1,000 fine-tuning steps with your proposed objective, starting from the released IRIS adapter, the editor completed 244 of 300 requests on protected photos, up from 1 of 300. On the clean versions of those same photos, successful edits fell to 247. So the attack reversed most of the protection with a slight utility cost to unprotected editing. We now state this limitation in Section 6 and report the full results in Appendix D.3.

This discovered weakness reminds us of the fine-tuning fragility of other types of parameter-level safeguards like concept erasure in text-to-image diffusion models (Suriyakumar et al. [1]), while Gao et al. [2] study whether meta-unlearning can make erased concepts harder to relearn. Future work could draw on meta-unlearning to make the opt-out harder to undo. Another possibility could be to introduce an IRIS-like objective during pre-training, so the model learns to respect the opt-out more deeply alongside its ordinary editing skills. As a basic mitigation, a model developer could decide to rely on a third party for the watermark embedding through a rate-limited API while editing stays local. A query limit could slow collection through that API. Each returned photo would still give the requester a pair, and we did not test this mitigation.


W1: Adaptive removal under the stated public-release assumptions is discussed but not evaluated.

We thank the reviewer for raising this point and would like to clarify that we do evaluate purification quantitatively, in Appendix Tables 11, 17, and 18. The revised main results now report the purification comparison. Appendix D also reports both adaptive attacks you proposed.

Table 11 tests watermark strength on 100 VetoBench-General photos. The separate comparison on 300 photos shows that stronger purification also substantially weakens the evaluated input-only defenses (Tables 17 and 18). Simply doubling the perturbation strength, with no retraining, IRIS blocks 94% of edits on FLUX.2 and 100% on Qwen after purification at t = 50 (Table 11). The perceptual cost remains modest: input LPIPS rises from 0.042 at native strength to 0.113 at doubled strength, which is still lower than the values reported for EditShield and PhotoGuard. Figure 7 illustrates a related option on one photo: concentrating the stronger signal in textured regions. Section 6 discusses why this needs testing on more images.


Thank you for suggesting both adversarial tests. They helped us draw a clearer boundary on the scope in the revised paper. We updated the paper PDF to report both attacks in Appendix D.2 (PGD, Table 16) and Appendix D.3 (reverse distillation) and updated the limitations section to reflect these new insights.

References

[1] Suriyakumar et al. Unstable Unlearning: The Hidden Risk of Concept Resurgence in Diffusion Models.

[2] Gao et al. Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Concepts. ICCV, 2025.

Dear Reviewer, as the discussion period is coming to an end, we would appreciate hearing from you. Thank you!

Reviewer_emB5OpenReview ↗

Response to Author Rebuttal

Thank you for addressing my concerns. I find the observations from the PGD attack quite intersting. Could you please tell me what objective was used in the PGD attack? Nevertheless, it would be an intersting future direction to invetigate futher why the protection is retained despite the disruption of the decoded message. Thanks for also investigating Q2 and reporting them. I hope these help the community for furthering this area of research and building practical solutions.

AuthorsOpenReview ↗

Clarification of PGD Objective

Thank you for taking another look and raising your score to weak accept. We really appreciate your suggestions and what we learned from these experiments.

WAM predicts two outputs at each spatial location: the presence of a watermark and 32 message bits. At ϵ=8/255\epsilon=8/255, our PGD attack minimized both via

\operatorname{mean}\left[\operatorname{softplus}(D_{\mathrm{pres}}(x))\right] + \operatorname{mean}\left[\max\left(0,(2m-1)D_{\mathrm{bits}}(x)\right)\right],$$ where $D_{\mathrm{pres}}$ and $D_{\mathrm{bits}}$ are the frozen decoder’s logits and $m$ is the known message. The first term encourages “no watermark.” The hinge term penalizes evidence for the correct bits until the decoder gets them wrong. Across 300 protected photos, mean presence probability fell from **99.98% to 0.17%**, and bit accuracy from **100% to 1.4%**, while IRIS still blocked **99.7%** of edits. We agree that understanding why protection survives this disruption is an interesting direction, and thank you for drawing our attention to it. Thank you again for the helpful discussion. If any concerns remain or anything would benefit from further clarification, we would be happy to address it.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Meta-Reviewer,

We appreciate the positive assessments of IRIS across the reviews and would like to begin by summarizing the strengths they identified:

  • Clarity & Effectiveness: All three reviewers (emB5, bChH, and Rq2g) found the paper clear and easy to follow, highlighted the simplicity of the method, and described effective protection with preserved ordinary editing under the evaluated conditions.

  • Extensive Evaluation: Reviewers emB5 and bChH highlighted the extensive empirical evaluation comparing against existing baselines across multiple models and detailed ablation studies supporting the approach.

  • Addresses Major Limitations: Reviewer emB5 additionally emphasized that IRIS conceptually addresses major limitations of earlier protection methods, including protecting one image across multiple cooperating editors and robustness to purification. They also highlighted that training does not require paired data and that the supplementary material is of high quality.

  • Practical Problem: Reviewer bChH emphasized that the paper addresses a meaningful problem and that selectively protecting images while preserving ordinary editing is important for practical usefulness.

  • Simplicity & Low Costs: Reviewer Rq2g additionally highlighted the low training cost and described the simple training objective as effective.

We posted detailed responses and new experiments on 13 November and followed up on 16 November. Following our response, Reviewer emB5 confirmed that their concerns were addressed and raised their rating from 2 to 3 (Weak Accept). For your convenience, we also summarize below how the new experiments and revisions address the concerns raised by Reviewers bChH and Rq2g.

  • Novelty beyond GuardDoor and Chen et al. (bChH, Rq2g): GuardDoor and Chen et al. established the cooperative protection setting. We built on that, but when we applied their approaches to modern FLUX.2, both failed to preserve editing on unprotected images. Each completed only 44.3% of requests on unprotected photos, compared with 88.3% for IRIS (Table R2; revised paper, Table 1). Our controlled ablations in Table R3 show that the choice of the response objective, i.e. what we teach the editing model to do for protected images, has an independent effect on selectivity when the other training choices are held fixed.

  • Clarity of zero-velocity regression (bChH, Rq2g): We expanded Section 3.2 to explain the key distinction and intuition: zero velocity provides the same target at every protected state and stalls denoising, while generating a fixed replacement image requires velocities that vary with the current state. Controlled comparisons now isolate the benefit of our velocity cooldown objective for protection while preserving ordinary editing (Table 3).

  • Purification (emB5, Rq2g): We clarified that purification weakens protection at the original watermark strength. However, doubling the strength without retraining blocked 94% of edits on FLUX.2 and 100% on Qwen under the tested DiffPure setting. This improvement comes with greater input distortion (Table 11).

We have already revised the submission and marked all additional content in blue. We respectfully ask that the final assessment consider these new results alongside the original reviews and the limitations now documented. To make the methodological contribution clearer, we would also be happy to revise the title to “IRIS: Imperceptible, Robust Image Sealing via Zero-Velocity Regression for Cooperative Anti-Edit Protection,” while retaining IRIS as the method name.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.