acceptodds
Withdrawn from ICLR 2027

I2VShield: Focused Adversarial Protection Against Harmful Image-to-Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern image-to-video (I2V) diffusion models turn any photograph into a high-fidelity video, making non-consensual deepfakes cheap to produce. A natural defense is adversarial protection: adding an imperceptible perturbation (under a small budget) to an image before publication so that generation from it fails. Yet existing methods inherit the text-to-image (T2I) recipe, which barely dents modern I2V backbones. We trace this failure to the recipe's two pillars, neither paying off under I2V. First, the 3D VAE is far harder to move than its 2D counterpart, so the same budget buys far less disruption here. Second, and more decisively, I2V's static conditioning leaves random-timestep averaging both unnecessary and self-diluting: the perturbation never enters the noisy latent, and we measure cross-timestep gradients to be near-orthogonal (cosine I2V vs T2I ). Probing where the condition acts, we locate its influence at the largest noise level, decaying monotonically thereafter, so averaging spends most of the budget where the condition barely responds. Together, these findings identify what decides protection strength: not which objective is attacked, but where along the timestep axis the budget is spent. Guided by this diagnosis, we propose I2VShield, which forgoes the VAE objective and commits the entire budget to the largest noise level, where the condition's influence peaks. Across 9 I2V models and 2 datasets, it lowers DINO similarity from to on average, against for the strongest, outperforming even at a smaller budget ( vs DINO). We further release HQFace-I2V, the first benchmark purpose-built for this setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Protection, Image-to-Video 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_y8QFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents I2VShield, a protection method against I2V manipulation of images based on adversarial examples. The paper addresses why existing protection methods against T2I models are not directly applicable to I2V models. First, the paper shows that optimization that disrupts the internal latent of VAE encoders is not applicable to I2V models. Second, T2I protection methods that use random time steps are not applicable to I2V models. To this end, the paper proposes I2VShield, which focuses adversarial optimization budgets on a single time step. Across various benchmarks, the proposed method outperforms existing SoTA protection methods against both T2I and I2V models.

Critical strengths
  1. The paper presents a neat analysis on why existing protection methods on T2I models do not transfer directly to I2V models. Instead of a simply argument, this analysis adds value on how protection against I2V models should be designed.
  2. The proposed method is evaluated against a wide range of I2V models and is also compared with various methods against both T2I and I2V.
  3. The proposed method heavily outperforms existing SoTA methods, including protections against both T2I and I2V models.
Critical weaknesses
  1. Technical contributions. Technical contributions of the paper are limited. Instead of optimizing the adversarial examples over all time steps, the paper uses a single time step. This is an approach taken by DiffusionGuard [1], which is a method targeted for T2I model yet uses a single time step.
  2. Size of evaluation set. 80 image-prompt pairs for evaluation seems too small. I2VGuard and Anti-I2V use 300 and 5,000 data pairs, respectively, which are far larger than the dataset size used in this paper.
  3. Bypassing protection. Since the proposed method focuses adversarial optimization only on a single time step, it brings a new level of manipulation, where the adversary can intentionally bypass the target time step τ∗\tau* used during optimization. For example, the adversary may use SDEdit to start sampling from a partially noised latent of the protected image, so that τ∗\tau* is never evaluated during generation. The paper needs evaluation on the robustness of protection against such scenario.

[1] Choi et al., DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Image Editing, ICLR 2025

Questions for authors
  1. Please clarify the technical contributions of the work, especially compared to DiffusionGuard.
  2. Would it be possible for an adversary to break the protection by bypassing the time step used for optimization? If so, would the proposed method remain robust?
Reviewer_VpoyOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes I2VShield, an adversarial image protection method against image-to-video generation. The method removes auxiliary VAE objectives and optimizes the difference between clean- and perturbed-condition predictions at the highest noise level. The authors motivate this choice through experiments on VAE reconstruction robustness, gradient alignment across timesteps, and attention to the conditioning frame. The paper also introduces HQFace-I2V and evaluates protection across multiple I2V models, perturbation budgets, prompts, and purification settings.

Critical strengths
  1. The method is simple and produces substantial improvements in the reported white-box experiments. Table 1 shows consistent gains across three backbones, two datasets, and five evaluation metrics.
  2. The evaluation extends beyond the main comparison to additional checkpoints, unseen prompts, perturbation budgets, purification, and generation settings. The small-budget results are particularly encouraging.
  3. Figure 5 provides useful evidence for timestep selection: with the proposed objective and optimizer fixed, attacking the highest noise level performs better than attacking the other tested individual timesteps.
  4. HQFace-I2V provides high-quality portrait inputs with multiple animation prompts. Evaluating on an additional public dataset helps reduce dependence on the newly introduced benchmark.
Critical weaknesses
  1. The experiments do not sufficiently separate timestep selection from the choice of attack objective. My main concern is the claim that protection strength is determined by where the attack budget is spent, rather than which objective is used. Figure 5 supports the highest noise level over the other tested individual timesteps for the proposed objective. However, this is different from establishing that timestep selection explains the improvement over existing methods. The main comparison changes both the objective and the timestep policy. Table 3 adds auxiliary terms to the proposed objective at the highest noise level, but does not compare alternative objectives under both highest-timestep and uniformly sampled-timestep optimization. It therefore does not establish that the choice of loss is unimportant or that uniformly sampled timesteps are inherently ineffective. The authors should compare the proposed prediction-divergence objective and a representative baseline objective under both timestep policies, using the same perturbation budget and reporting comparable optimization costs. A limited experiment would be sufficient to test this explanation. Without such a comparison, the conclusion should be restricted to the demonstrated effectiveness of the complete I2VShield configuration.

  2. The baseline adaptations need further validation, particularly for Anti-I2V. Appendix E.3 provides useful implementation details, including the adaptation of T2I objectives to image conditioning, the shared bank of precomputed clean trajectories, and the original iteration counts. It also explicitly states that Anti-I2V’s CLIP and LPIPS auxiliary terms were removed. These choices are disclosed, but their effect on baseline strength has not been established. There are additional differences that need clarification. The original Anti-I2V paper uses dual-space perturbation updates in the Lab∗L^{}a^{}b^{*} and DCT domains, together with layer-wise representation anchoring in the VAE and denoising network. The present manuscript describes pixel-space sign-gradient optimization and expresses the IRA term as a penalty on the final conditioning latent. It is unclear whether the original perturbation parameterization and layer-wise anchoring are retained, replaced, or approximated. These differences do not establish that the implementation is incorrect, but they matter because Anti-I2V is the strongest reported baseline and is used to support several of the paper’s main comparisons. An ℓ∞\ell_\infty constraint also does not, by itself, establish that removing perceptual auxiliary losses leaves an attack’s behavior unchanged. The authors should list which components of Anti-I2V were retained or modified, explain the correspondence between the original and implemented objectives, and validate the important adaptations under the common perturbation constraint. A faithful reproduction or a clearly justified and adequately tuned adaptation is needed to support the claimed advantage over prior methods.

  3. The gradient-alignment diagnostic is not sufficiently specified, and its interpretation is too strong. The additional models and noise averaging in Appendix C.2 strengthen the observation that the tested I2V gradients share less directional alignment than the tested T2I gradients. However, the exact diagnostic objective A\mathcal A is not specified. This is important because the gradients are reportedly evaluated at the original image. If A\mathcal A is the squared clean-versus-perturbed prediction difference used by I2VShield in Equation (7), then ∇δA(0)=0\nabla_\delta \mathcal A(0)=0 under matched inputs and noise. A different diagnostic objective may have been used, but it needs to be stated explicitly before the reported cosine similarities can be interpreted. Furthermore, low pairwise cosine similarity indicates limited directional agreement; it does not by itself establish that stochastic timestep optimization produces weaker final attacks. It also does not identify the highest noise level as the best single timestep. The authors should provide the exact loss, evaluation point, and gradient paths used in this diagnostic. The interpretation should distinguish the measured gradient geometry from claims about optimization performance, which require a direct comparison such as the one described in the first weakness.

  4. The VAE experiment supports reconstruction robustness more directly than encoder insensitivity. The VAE analysis is useful, and Appendix C.1 includes both an F=1F=1 capacity control and several reconstruction metrics. Nevertheless, the reported measurements compare reconstructed images with the original input. They do not directly establish that the encoder representation changes less, since the decoder can also affect how latent changes appear in the reconstruction. The experiment therefore supports a narrower conclusion: the tested 3D autoencoders retain more reconstruction similarity under the specified attack. It does not fully establish that perturbations are absorbed by the encoder or that VAE-side objectives generally have little value for I2V protection.

  5. Attention to the conditioning frame is not a direct measure of adversarial sensitivity. The attention measurements on Wan provide a plausible explanation for prioritizing high-noise timesteps, and their trend is consistent with the single-timestep ablation. However, a larger attention share does not necessarily imply a larger input Jacobian, a stronger attack gradient, or a larger effect on the final generated video. Similarly, the fact that the image condition is static does not by itself establish that optimizing across multiple timesteps is unnecessary or ineffective. The model’s response to that condition can still change across the trajectory. The authors should present the attention results as supporting observations rather than as proof that the highest noise level captures all useful conditioning influence. If the stronger sensitivity claim is retained, it needs a direct sensitivity or intervention experiment. This point concerns the explanation for timestep selection and should be considered together with the first weakness.

Questions for authors
  1. Timestep selection versus objective design. Could you compare the proposed prediction-divergence objective and a representative baseline objective under both highest-noise and uniformly sampled timestep optimization, using the same perturbation budget and reporting optimization costs, to determine whether timestep selection explains the gains over existing methods?

  2. Anti-I2V reproduction. Could you clarify whether Anti-I2V’s dual-space updates in the Lab∗L^{}a^{}b^{*} and DCT domains and its layer-wise IRA losses were retained, and provide evidence that the adaptations—including the removal of CLIP and LPIPS terms—do not materially weaken this baseline?

  3. Gradient-alignment diagnostic. Could you specify the exact diagnostic objective A\mathcal A, perturbation evaluation point, and gradient paths in Appendix C.2, particularly given that the squared prediction-difference objective in Equation (7) would yield ∇δA(0)=0\nabla_\delta\mathcal A(0)=0 under matched inputs and noise, and explain how the measured alignment relates to the performance of stochastic timestep optimization? I am open to revising my rating if the authors adequately address the concerns above through clarification and, where needed, additional experimental evidence.

  4. VAE robustness interpretation. Could you report appropriately normalized latent deviations and their relationship to downstream generation effects to support the claim of encoder insensitivity, or clarify whether the conclusion should be limited to the observed reconstruction robustness of the tested autoencoders?

  5. Attention and adversarial sensitivity. Could you support the proposed link between conditioning-frame attention and attack effectiveness with direct sensitivity or intervention measurements, or qualify the attention results as supporting observations rather than evidence that the highest noise level captures all useful conditioning influence?

I would be happy to reconsider my score if the authors can satisfactorily address the concerns above.

Reviewer_Wb1KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Adversarial protection adds an imperceptible perturbation to a photo to disrupt image-to-video (I2V) generation, but existing methods reuse the text-to-image recipe of a VAE-encoder attack and a random-timestep diffusion loss, and barely degrade modern I2V models. To address this, the paper proposes I2VShield, a simple attack that maximizes the prediction change at a single timestep. Specifically, it removes the VAE objective because the 3D VAE absorbs most of the perturbation, and it attacks only the largest noise level because cross-timestep gradients are nearly orthogonal and the image condition matters most there, claiming that this collapses the generated videos. Experiments on several I2V models and datasets, including the new HQFace-I2V benchmark, show that it outperforms existing baselines and stays effective under unseen prompts and smaller budgets. By validating its diagnosis, the paper effectively shows that protection strength depends on where the budget is spent along the timestep axis. However, it assumes white-box access, transfers poorly across models, and is weakened by strong purification.

Critical strengths
  • S1: The paper systematically diagnoses why existing protections fail on I2V models and derives a simple, easy-to-understand method from this diagnosis. It shows the difference between T2I and I2V through VAE robustness, cross-timestep gradient alignment, and attention analyses, and it checks these findings on several models and metrics. Since each design choice follows directly from the diagnosis, the method uses only a single objective at the first timestep, without a target image, a pre-generated clean video, or per-model tuning.
  • S2: The proposed method shows large and consistent gains across many models and settings, which makes its effect convincing. On three backbones with different architectures and two datasets, it clearly outperforms six baselines, including two recent I2V-specific methods, on all metrics. The gains also hold on six additional models, under unseen prompts, and with a much smaller perturbation budget, so the result is not limited to a single setting.
Critical weaknesses
  • W1: The main contribution is not clearly established. The core design, attacking only the highest-noise step of a static image condition, closely follows prior protections for inpainting models (DDD, IJCAI 2024; DiffusionGuard, ICLR 2025), which are not cited, and the finding that the condition matters most early in sampling is also known (e.g., eDiff-I). The claim that would set the paper apart, that protection strength depends on where the budget is spent rather than on which objective is attacked, is not isolated, because I2VShield also replaces the denoising loss with a new prediction-divergence objective and no experiment crosses the objective with the timestep strategy. As a result, it is unclear whether the gains come from the timestep choice, the new objective, or how the baselines were implemented.
  • W2: The protection is not shown to hold under a realistic threat model. All results assume white-box access to the exact target model and its generation settings, while the photo owner cannot know which model or sampler an attacker will use, and the perturbation transfers poorly across backbones (App. F.4). Because the method relies on the first sampling step, its effect weakens when the attacker simply uses more sampling steps or a few-step distilled model (Tab. 7–8), and strong purification removes it (App. F.2). The main experiments also use shorter settings than the official defaults, and no defense-aware adaptive attack is evaluated, so the practical value of the protection remains uncertain.
Questions for authors
  • Q1: How do the results change when the objective and the timestep strategy are separated? Specifically, what are the results of (i) the proposed prediction-divergence objective averaged over uniformly sampled timesteps, (ii) the AdvDM and Anti-I2V objectives evaluated only at τ* = 1, and (iii) a DiffusionGuard-style objective (the norm of the first-step prediction) adapted to I2V? Without these results, it is unclear whether the gain comes from where the budget is spent or from the new objective.
  • Q2: DDD and DiffusionGuard already focus the budget on the earliest denoising steps for image inpainting models with a static condition. Which part of the finding is specific to I2V? Do the same diagnostics (cross-timestep gradient alignment and attention to the condition) show a different pattern on an inpainting model than on I2V models, and if not, how is the contribution positioned against these works?
  • Q3: How well does the protection hold under choices the attacker can make at no cost to video quality, such as the official default settings (e.g., 81 frames and 40–50 steps for Wan2.1), other solvers, few-step distilled models, and a defense-aware attacker who lightly re-noises or purifies the condition image before generation? Does optimizing the perturbation over an ensemble of backbones reduce the weak cross-model transfer?

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

Related papers

Discussion (0)

Sign in to comment.