acceptodds
Withdrawn from ICLR 2027

Immune2V: Image Immunization Against Dual-Stream I2V Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image-to-video (I2V) generation has the potential for societal harm because it enables the unauthorized animation of static images to create realistic deepfakes. Recent I2V immunization methods protect reference images by introducing imperceptible adversarial perturbations, but remain fragmented across specific architectures and conditioning mechanisms. In this paper, we systematically analyze how reference images participate in modern I2V generation and identify a common dual-stream conditioning paradigm: a that provides a visual anchor, and a that interacts with other semantic conditions (e.g. text prompt) throughout the generation process. Motivated by the need to jointly disrupt these two streams, we propose framework which maximizes the deviation between perturbed and clean structural conditioning signals across distinct conditioning mechanisms, while aligning intermediate generative representations with collapse-inducing trajectories across diverse plausible prompts to counteract semantic guidance. Experiments demonstrate that Immune2V provides stronger and more persistent protection than existing baselines, consistently disrupting I2V generation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Attacks, Image Immunization, I2V Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZQhUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Immune2V, which adds a bounded adversarial perturbation to an image before release so that videos generated from it by image-to-video (I2V) models are distorted, as a defense against deepfakes made from personal photos. Treating the reference image as entering generation through a structural and a semantic stream, it combines a structural loss that pushes the perturbed image's structural condition away from the clean one (with a temporal coverage term on Wan 2.1) and a semantic loss that pulls early DiT features under aligned prompts toward those obtained with collapse-inducing prompts. Experiments on Wan 2.1 and HunyuanVideo-I2V compare against four baselines and include a loss ablation, input transformations, and cross-model transfer.

Critical strengths

S1. Optimizing and evaluating on 13B to 14B open-source DiT models (Wan 2.1 and HunyuanVideo-I2V) keeps the setting close to a realistic threat, and perturbations optimized on Wan 2.1 still lower all five metrics on CogVideoX and DynamiCrafter without re-optimization.

S2. The structural attack is built on a concrete observation about Wan 2.1. Because the reference image is zero-padded into a pseudo-video before the causal VAE, an image-level perturbation mostly affects the first few temporal latent slices, and the temporal coverage term is designed to spread the disruption across all of them. This is the part of the paper I found most informative.

Critical weaknesses

W1. The evaluation does not adequately support the claimed protection Sec. 3.1 defines success as "severe and persistent degradation across the entire generated video", but Table 1 reports only five averaged metrics. On Wan 2.1, SC drops from 95.85 to 90.34, which does not show how many videos actually become unusable. There is no per-sample failure criterion, human evaluation, or variance. The paper is motivated by deepfakes made from personal photos, yet the 50-image face subset has no identity-related evaluation, although Anti-I2V, one of the baselines, measures identity preservation with ArcFace. The robustness claim in Sec. 5.2 is also stronger than Table 3 shows. After JPEG compression, about two-thirds of the SC drop is lost, and results under diffusion-based purification are not reported.

W2. The distinction from Anti-I2V drawn in Sec. 2 does not hold, and the ablation isolates neither the specific design choices nor the two streams Sec. 2 states that existing I2V immunization methods "mostly attack the old UNet denoising backbone", but Anti-I2V already reports results on DiT models such as CogVideoX-5B and OpenSora and also targets VAE and denoiser representations, which is where the two losses here operate. What remains specific to this paper is the collapse-prompt feature target, the expectation over prompts, and the temporal coverage term. Table 2 supports combining the two objectives but does not isolate the contribution of these choices. Nor does it necessarily isolate the two streams. Eq. (8) uses hadvh^{\text{adv}} computed from the whole IadvI_{\text{adv}}, so unless the gradient through the structural condition is stopped, LSem\mathcal{L}_{\text{Sem}} also acts on the structural stream. The motivating analysis is also underspecified. Fig. 3 does not state which image-level attack is used, and Fig. 4 does not specify how the semantic distance is computed.

Questions for authors
  1. Could you state the threat model, in particular which generation settings the attacker is assumed to control? On Wan 2.1, which CFG scale is used at test time, and are the random seeds different from those used to build the collapse trajectories? If possible, how much protection remains with different step counts, CFG scales, or prompt styles?
  2. What parameters are used for the three transformations in Table 3, and what are the five metrics when the same transformations are applied to clean images?
  3. How is the semantic distance in Fig. 4 computed, and which attack produces the "Struct. Attacked Image"? Fig. 5 draws the backward path of LSem\mathcal{L}_{\text{Sem}} only through the semantic encoder. Is the gradient through the structural condition detached in the implementation?
  4. For each baseline adapted to the two models, which modules are attacked, and what are the iteration count and per-image runtime? On Wan 2.1, PhotoGuard's PIC (84.88) is close to that of the clean input (85.26), while LStruct\mathcal{L}_{\text{Struct}} alone, which also attacks the encoder latent, reaches 82.71. Which encoder and target latent does PhotoGuard use? What PSNR, SSIM, or LPIPS do the immunized images of each method reach?
  5. The paper states that the code will be released upon acceptance. Could you share an anonymized implementation during the discussion, in particular for the four reimplemented baselines?
Reviewer_LFtDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies image immunization for image-to-video (I2V) generation, which protects reference images with imperceptible adversarial perturbations so they cannot be animated into realistic fake videos. The authors observe that modern I2V models condition on the reference image through a dual-stream paradigm: a structural stream and a semantic stream. They propose Immune2V, which jointly maximizes the deviation between the perturbed and clean structural conditions and steers the semantic conditioning toward collapse-inducing trajectories. Experiments on the authors’ benchmark show stronger protection than existing baselines.

Critical strengths
  • The research problem is important and practically relevant.
  • The structural and semantic attack objectives are intuitive and simple.
Critical weaknesses
  • The evaluation should cover a broader set of I2V generators. In particular, I recommend evaluating transfer to closed-source, commercial I2V systems. Users would need protection against these in real-world deployment, and white-box optimization is not possible for them. -Immune2V is itself an adversarial perturbation, yet the paper does not test its robustness against attacks designed to remove or neutralize it (e.g., adaptive attacks to Immune2V or diffusion-based purification)
  • The robustness analysis should be more thorough. The current study covers only Gaussian filtering, JPEG compression, and resizing, without specifying their strengths. How do the baselines perform under the same transformations? Without this comparison, it is unclear whether Immune2V is more robust to perturbation removal than existing methods
  • The paper states that the method “does not rely on the specific choice of encoders” (Appendix A.2), but no experiment supports this. An ablation across different visual, semantic, or text encoders would substantiate the claim.
Questions for authors
  • Could the authors provide experiments with different visual or text encoders to support this claim?
  • How does Immune2V compare to the baselines under the same input transformations?
  • How does Immune2V perform against closed-source I2V generators?
Reviewer_Sj6yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies image protection against unauthorized image-to-video generation. The authors describe structural and semantic conditioning streams and propose Immune2V to attack both. A structural loss increases reference-feature deviation, while a semantic loss matches features from generations driven by conflicting prompts. Experiments on Wan 2.1 and HunyuanVideo-I2V, with transfer tests on other models, show stronger average generation degradation than the evaluated baselines.

Critical strengths
  1. The method is tested on two recent video generators and compared with recent defenses. The ablations show that combining the two losses causes greater generation degradation.

  2. The authors test four prompts that were not used during optimization. They also examine compression, resizing and transfer to other models.

Critical weaknesses
  1. The collapse prompts are chosen because they conflict with the image content. The paper does not verify that their videos collapse rather than show a different but coherent scene.

  2. Average video scores can decrease while an animation remains usable. The paper should report how often videos actually become unusable or lose the protected identity.

  3. The claim that recent methods mainly target old U-Net models is inaccurate. Anti-I2V and Anti-Prompt already study modern architectures.

  4. Figure 4 shows changes in distance between generation trajectories. This alone does not establish that semantic guidance repairs structural corruption.

  5. The perturbation budget alone does not establish that the protected images look unchanged. Input image quality should also be measured.

Questions for authors
  1. Do verified collapse targets outperform conflicting targets that still generate coherent videos?

  2. On the existing test prompts, how often are outputs independently judged unusable or unable to preserve identity?

  3. Could the authors isolate semantic correction from ordinary changes during denoising?

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

Related papers

Discussion (0)

Sign in to comment.