acceptodds
Under review as a conference paper at ICLR 2027

BadWorld: Adversarial Attacks on World Models

Abstract

Visual world models (VWMs) synthesize interactive, action-conditioned rollouts from a single context image. However, it remains an open question how robust these models are to adversarial perturbations. Standard adversarial attacks fail to assess this vulnerability because attackers lack ground-truth future videos and cannot predict subsequent user controls. We introduce BadWorld, a label-free adversarial framework tailored for autoregressive VWMs that systematically overcomes both constraints. First, to bypass the need for future supervision, we propose a self-supervised velocity attack that directly disrupts the early denoising dynamics of the model. Second, to ensure the attack generalizes across unpredictable user actions, we formulate a trajectory-adaptive bi-level optimization that actively mines hard control sequences to forge control-agnostic perturbations. Evaluated on representative VWMs with continuous and discrete controls, BadWorld exposes severe structural fragility. Visually indistinguishable adversarial images reliably trigger catastrophic degradation in future rollouts, leading to incomplete denoising, structural collapse, and control inconsistency. Crucially, this visual corruption propagates to trajectory planning in autonomous driving, causing severely unsafe decisions. These findings reveal critical risks for deploying VWMs in safety-critical systems while highlighting a practical mechanism for privacy protection.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial attack, autoregressive video models, world models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DSmPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the vulnerability of interactive visual world models, where even a slight perturbation to the initial input can lead to eventual failure. It proposes a label-free adversarial method to achieve consistent attacks under diverse future controls. Importantly, experiments in autonomous driving scenarios show that such attacks can lead to severe safety consequences.

Critical strengths
  1. This paper is well written and easy to follow.
  2. The studied problem is important, i.e., robustness in interactive visual world models.
Critical weaknesses
  1. The proposed method assumes that the attacker has full model access, which is unrealistic in real-world deployment scenarios. When the model is inaccessible, it is suggested that the authors discuss the possible transferability of the method to strengthen its value.
  2. The authors claim that a major challenge for the attack is obtaining the future ground-truth video. Considering that the ultimate goal of the attack is to corrupt the generated results, why cannot Equation 4 directly align Gθ(x + δ, τ1:K) with random noise or with the results under the opposite control? Therefore, the ground-truth video can be easily bypassed rather than being a key challenge.
  3. The authors claim that the perturbations are imperceptible. However, the evaluation only considers human-level perception, and it remains unclear whether the perturbations introduce detectable statistical artifacts. A simple adversarial detector or perceptual analysis could help quantify whether the perturbations are genuinely imperceptible rather than merely visually subtle.
Questions for authors

Please refer to the issues raised in the Weaknesses section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer DSmP

We thank the reviewer for recognizing the clarity of our presentation and the importance of robustness in visual world models. We will address your concerns in the following response.

>>> W1 Transferability Without Target-Model Access

>>> A1 We appreciate this practical suggestion. We completely agree that transferability is important when the target model is inaccessible. To prove this, we present the existing black-box transfer results from Appendix A.3 and add another cross-model evaluation.

  • Existing evidence: Appendix A.3 evaluates transfer between Matrix-Game 2.0 and Astra. Adversarial images are resized to the target model’s input resolution before inference.
Training → TestInputBGC ↓Aesth. ↓Imaging ↓
Matrix-Game → AstraClean0.8780.4920.553
Matrix-Game → AstraAdversarial0.8460.4660.515
Astra → Matrix-GameClean0.9550.4880.659
Astra → Matrix-GameAdversarial0.9540.4850.645

The Matrix-Game → Astra transfer reduces Background Consistency from 0.878 to 0.846 and Imaging Quality from 0.553 to 0.515 without re-optimizing on Astra. This confirms that BadWorld perturbations remain effective on an unseen target model.

  • Additional evaluation: We further evaluate 30 images by applying perturbations learned on Matrix-Game GTA directly to Matrix-Game Universal, without target-model re-optimization.
InputBGC ↓Aesth. ↓Imaging ↓CLIP-I ↓
Clean0.9600.5360.6970.885
Transferred attack0.9410.5210.6090.803

The transferred attacks degrade all four metrics, reducing Imaging Quality from 0.697 to 0.609 and CLIP-I from 0.885 to 0.803. The additional evaluation further supports transferability beyond the white-box model used for optimization.

>>> W2 Targeted Objectives Still Need a Rollout Proxy

>>> A2 Thanks for raising this thoughtful question! We believe there is a misunderstanding here. A random-noise or opposite-control video can provide target supervision for a targeted attack, but it cannot replace the video latent required as denoising input.

Specifically, under a training-style diffusion denoiser, we must first construct a noisy latent by adding noise to a clean video latent. However, in our setting, only the context image is available; this ground-truth video latent is missed. Changing the supervision target therefore changes only the loss and still leaves us without a valid denoising input or autoregressive history condition. BadWorld directly solves this by using repeated context frames as a surrogate history and self-predicted velocity as label-free supervision.

>>> W3 Imperceptibility

>>> A3 Thanks! We agree that visual subtlety alone does not rule out detectable statistical artifacts. To first quantify the perturbation magnitude, as stated in line 413, our default perturbation budget is 0.05 in the normalized pixel range [−1,1][-1,1]. This corresponds to 0.0250.025 in [0,1][0,1], or approximately 6.375/2556.375/255.

  1. Perceptual analysis: We report PSNR, SSIM, and LPIPS between clean and adversarial input images for attacks on Astra. The table reports mean ± standard deviation.
MethodPSNR ↑SSIM ↑LPIPS ↓
BadWorld34.981 ± 0.3350.9037 ± 0.03080.1352 ± 0.0566

(Arrows indicate better preservation of the original image.)

These results show that BadWorld introduces only small perceptual changes and maintains high fidelity to the clean input.

  1. Statistical detection. We additionally evaluate a simple detector that learns normal noise and frequency statistics from clean images, then scores unseen images for anomalies. We test it on 30 clean/adversarial pairs using five-fold cross-validation repeated 20 times, with no overlap between training and test images. It achieves AUROC = 0.603 ± 0.010, where 0.5 indicates random guessing and 1.0 indicates perfect detection. This suggests that, although statistical differences may exist, the detector has limited ability to distinguish adversarial images from clean ones.
Reviewer_NYR3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether small perturbations added to an initial image can distort the videos produced by a video generation model. The proposed method targets the denoising process and optimizes perturbations against control sequences that are most resistant to the attack. Although the attacks are clearly effective, the paper does not convincingly justify the significance of the work.

Critical strengths

The proposed approach has two main strengths:

  1. It effectively applies adversarial attack principles and achieves strong attack performance.

  2. It accounts for the different control trajectories that a world model may receive and remains effective across these trajectories.

Critical weaknesses

The main weaknesses are:

  1. It is difficult to identify realistic application scenarios for the attack. For models used purely for content generation, I do not consider supplying an adversarial image simply to cause generation to fail a meaningful attack. For WAMs such as Epona, the inputs originate from physical scenes. If an adversarial attack cannot be deployed in real-world settings, it has no practical value. The paper lacks experiments addressing this issue.

  2. The authors should report Epona’s collision rate to demonstrate the effectiveness of the attack.

Questions for authors

Same as weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer NYR3

We thank the reviewer for recognizing the effectiveness of our adversarial attack and its robustness across diverse control trajectories. We will address your concerns in the following response.

>>> W1 Application Scenarios and Practical Value

>>> A1 For sure, we fully agree that practical value matters, and in fact, it has been a key motivation behind our work. We believe BadWorld goes beyond simply disrupting video generation. It helps us understand the vulnerabilities of world models and may help to reveal identify potential safety risks in real-world decision-making. We are happy to clarify these applications and share new experimental results.

1. Practical Value for Visual World Models

  • a. Exposing fragility before deployment: Our goal is not simply to ruin generated videos. Instead, BadWorld exposes vulnerabilities in visual world models. Tiny, imperceptible perturbations can completely break both visual predictions and underlying physical dynamics. If robotics and autonomous driving systems are going to rely on these models, their reliability would be a super important topic before they cause real-world harm.
  • b. Protecting private visual assets: Beyond safety testing, BadWorld could also help people protect their own images. Today, world models can turn a single personal photo or copyrighted image into an interactive video scene, potentially without the owner's permission. Our attack offers a simple way to discourage such unauthorized generation by adding imperceptible perturbations before sharing images online.

2. Practical Safety Risks for WAMs such as Epona

  • a. Digital-stream interception: Even though our method currently attacks digital images, the threat is still realistic. For example, in cloud-assisted driving, remote teleoperation, or V2X systems, attackers may inject perturbations into camera inputs transmitted over networks. This could mislead world models without modifying the physical environment.
  • b. Localized-patch attack in the physical scene: We also agree that digital access may not always be available. An alternative is to place a printed adversarial patch within the camera's view, as demonstrated in prior work [1].

To further explore this possibility, we add a localized-patch attack experiment, restricting perturbations to 5% of the image while keeping the same BadWorld objective.

Results. We evaluate 50 clips covering diverse driving scenes and report task-level measurements over a 10-second horizon:

InputFVD₁₀ ↑FID ↑Avg.L2 (1/2/3s) ↑†CR≤10s ↑
Clean621.4037.251.3110%
Localized patch2884.83213.312.2530%

(† denotes newly added metrics, will be illustrated in W2, A2. Arrows indicate greater attack effectiveness.)

A patch covering less than 5% of one historical frame triples the measured collision rate from 10% to 30%. These results demonstrate that a localized patch placed on the road can still cause substantial driving-safety failures, establishing world-model robustness as a critical requirement. Reviewers MDJc and Reviewers DSmP also highlight our studied problem as timely and practically important.

>>> W2 Collision-Rate Evaluation

>>> A2 We truly appreciate this valuable suggestion. Following Epona’s collision-evaluation principle, we check whether predicted ego trajectories intersect surrounding objects. We report collisions for both initial 3s plans and autoregressive rollouts over 5s and 10s, and expand the evaluation to 50 clips covering different driving scenes.

InputFVD₁₀ ↑FID ↑Avg. L2 (1/2/3s) ↑†Plan CR≤3s ↑†Rollout CR≤5s ↑†Rollout CR≤10s ↑
Clean621.4037.251.310%2%10%
Adversarial3705.00363.653.414%24%40%

(† denotes newly added metrics. CR is the percentage of clips with at least one collision within the specified horizon.)

BadWorld increases collision rates for both initial plans (0% → 4%) and autoregressive rollouts (2% → 24% at 5s; 10% → 40% at 10s). This provides additional evidence that the attack compromises driving safety beyond degrading video quality. We will add these results and the evaluation protocol to the revised manuscript.


Reference:

[1] Wang, Taowen, et al. "Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics." ICCV, 2025.

Thank you for the additional experiments, which address most of my concerns; however, my concern about physical-world applicability remains unresolved, so I maintain my rating.

Dear Reviewer NYR3,

Thank you for your thoughtful follow-up! We really appreciate your careful consideration of our applicability.

To make the physical-world applicability more concrete and intuitive, we provide an anonymous project page (https://anonymouspage123.github.io/badworld-page/) with 5 matched Epona rollouts in autonomous driving scenario. These videos show that a patch perturbation on the road region leads the future predictions to be completely corrupted. More importantly, this leads to unsafe driving trajectory, and the car shows more collisions with nearby vehicles.

As an extension demonstration, we have not tested a printed patch on real roads. However, based on these videos and quantitative results, we believe that this attack could pose serious safety risks when world-model predictions inform driving decisions. Beyond autonomous driving, BadWorld also exposes broader world-model vulnerabilities and offers a way to protect private visual assets, as explained in W1/A1.

We hope these results help make the practical significance of our work clearer. If you have any remaining concerns, please do let us know!

Best regards,

The authors

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer NYR3,

Thank you for your valuable feedback and efforts during the review process!

As the discussion period draws to a close (less than 9 hours), we wanted to follow up on your question about physical-world applicability. We have responded to this point and would be grateful if you could take a look at our reply and the revised manuscript, where we have marked the changes in blue. If any aspect of our response needs further clarification, we would be happy to address it before the discussion closes.

Thank you again for your time and for the opportunity to discuss our work.

Best,

The authors

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

This paper proposes BADWORLD, a label-free white-box attack on autoregressive visual world models. It optimizes early-stage denoising velocity using a repeated-context history proxy and employs CMA-ES for hard-trajectory mining. Experiments on Astra, Matrix-Game 2.0, and Epona demonstrate degraded video generation and increased trajectory prediction errors.

Critical strengths

1.Relevant problem. Studying adversarial robustness in interactive, autoregressive world models is timely and practically important.

2.Efficient formulation. The velocity-based surrogate avoids future-video supervision and costly backpropagation through full rollouts.

3.Broad evaluation settings. Experiments cover continuous and discrete controls, with additional evidence of downstream trajectory prediction degradation on Epona.

Critical weaknesses

Major Weaknesses

1.Insufficient adversarial baselines. The main external baseline is random noise matched by mean and variance (Sec. 5.1). Without comparisons to rollout-discrepancy PGD or adapted I2V attacks, the advantage of the proposed velocity objective remains unclear.

2.Unvalidated surrogate objective. The attack uses early denoising timesteps (950–1000) and repeated context frames as a history proxy. Table 4 shows their empirical effectiveness but does not establish whether the surrogate reliably reflects full-rollout degradation under realistic histories.

3.Lack of evidence for autoregressive error amplification. The paper attributes attack effectiveness to errors accumulating through autoregressive feedback. However, Appendix A.2 only examines velocity scaling, without isolating the contribution of history feedback from general denoising instability.

4.Missing control-specific evaluation. VBench/VBench++, CLIP-I, and MEt3R primarily measure visual or geometric degradation, not whether generated rollouts follow the commanded actions. Thus, the claimed failure of world dynamics is not directly established.

5.Limited benefit of bi-level optimization. Improvements over Velocity-Min are modest and sometimes negative (Table 1: Imaging Quality 0.513 vs. 0.524). The bi-level method is evaluated only on Astra, without systematic held-out control tests or compute-matched comparisons with simpler trajectory sampling.

6.Overstated driving safety claims. The Epona study uses only 20 recorded clips and reports open-loop prediction errors. Without closed-loop planning or collision evaluation, the claim of "severely unsafe driving decisions" is not sufficiently supported.

Additional Weaknesses

1.Limited statistical reliability. Results lack confidence intervals and repeated-seed analyses. This is particularly concerning for the small bi-level improvements and the 10-image supplementary evaluations.

2.Unquantified imperceptibility. The default perturbation budget is 0.05, but no PSNR, SSIM, LPIPS, or clear pixel-range convention is provided to substantiate the imperceptibility claim.

3.Limited transferability and preprocessing robustness. Table 9 shows strongly asymmetric cross-model transfer, while Table 8 indicates substantial degradation of attack effectiveness after JPEG compression. Both evaluations use only 10 images.

Questions for authors

1.How does BADWORLD compare with rollout-discrepancy PGD or adapted I2V attacks under matched perturbation and compute budgets? (W1)

2.Does the early-timestep/repeated-history surrogate correlate with actual rollout degradation? Can history-reset experiments isolate autoregressive error amplification? (W2–W3)

3.Can the authors provide control-adherence or state-transition metrics beyond video quality? (W4)

4.Does CMA-ES improve performance on unseen controls compared with simpler trajectory sampling under matched compute? (W5)

5.Can the driving safety claims be supported by task-level evaluation or appropriately qualified? (W6)

6 comments
AuthorsOpenReview ↗

Response to Reviewer MDJc [1/4]

We thank the reviewer for recognizing the importance of this problem, the efficiency of our formulation, and the breadth of our evaluation. We will address your concerns in the following response.

>>> W1&Q1 Additional Baselines

>>> A1 Nice suggestion! We agree that stronger baselines are needed to clarify the benefit of our method. Following the suggestion, we explain our rollout-discrepancy surrogate and additionally evaluate two representative I2V attacks, I2VGuard [1] and Immune2V [2].

1. Rollout-discrepancy PGD. Directly optimizing Eq.(4) is not feasible in our autoregressive setting: it requires backpropagating through the full rollout, including all denoising steps and autoregressive chunks. We therefore use Drift-Max as a tractable surrogate for rollout discrepancy, which maximizes the discrepancy between the predicted velocity and a static reference velocity. It is already reported in Tables 1–2.

2. I2V attacks. We implement two representative I2V attacks: I2VGuard[1], which attacks temporal-attention features, and Immune2V[2], which jointly attacks spatial-temporal and semantic conditioning streams. Both use the same ϵ=0.05\epsilon=0.05 budget and 300 optimization steps as BadWorld. We summarize the results as follows:

MethodI2V-BG ↓BGC ↓Aesth. ↓Imaging ↓CLIP-I ↓MEt3R ↑
Clean0.9780.9150.5010.6900.8130.151
Drift-Max0.9380.8660.4660.5640.7500.207
Velocity-Min0.9300.8450.4050.5130.7140.264
I2VGuard0.9610.8900.4570.6580.7540.150
Immune2V0.9650.8960.4730.6720.7580.154

These results show that our simple Velocity-Min objective produces the largest degradation across all primary quality metrics on Astra, outperforming Drift-Max, I2VGuard, and Immune2V on Astra. We will add these comparisons and settings to the revised manuscript.

>>> W2&Q2 Validating the Surrogate Design

>>> A2 We appreciate this thoughtful concern. To test whether our surrogate reliably reflects full-rollout degradation, we revisit Table 4 to separate the effects of timestep selection and history construction. We also add Full-range + Self-Rollout, which uses full-range timesteps and histories generated by the model itself during attack. We report both attack effectiveness and runtime.

Attack Setting.

  • Timestep: Early samples from 950–1000; Full-range samples uniformly from 0–1000.
  • History: Context-Repeat constructs a history proxy by duplicating the input image. Self-Rollout generates four videos under different trajectories every 50 attack steps and samples histories from the refreshed pool, as described in line 422.
TimestepHistoryI2V-BG ↓BGC ↓Aesth. ↓Imaging ↓Time (min/img)
EarlyContext-Repeat0.9300.8450.4050.51311
Full-rangeContext-Repeat0.9760.9130.5140.71511
EarlySelf-Rollout0.9320.8430.4100.51460
Full-rangeSelf-Rollout0.9720.9150.5030.71260

(The first three rows are from Table 4.)

Comparing row 1 (Ours) with row 4 (Full-range + Self-Rollout), we achieve comparable degradation, and even stronger degradation on aesthetic and Imaging scores.

Notably, Context-Repeat achieves this performance at only 1/5 of the time cost. Early-timestep optimization is also consistently more effective under both history settings. Together, these results demonstrate that our surrogate provides an effective and efficient way to optimize perturbations.

AuthorsOpenReview ↗

Response to Reviewer MDJc [2/4]

>>> W3&Q2 Isolating Autoregressive Error Amplification

>>> A3 Thanks for this valuable suggestion. As suggested, we add a controlled history-reset experiment to test whether autoregressive feedback amplifies errors beyond denoising instability.

Setting. Following Appendix A.2, we scale the predicted velocity by 0.9 in every chunk. The two conditions use the same velocity scaling; only the history differs.

  • Self rollout: Each subsequent chunk uses the history generated by the velocity-scaled rollout.
  • History reset: From the second chunk onward, each chunk uses the corresponding clean history.

In that sense, both settings introduce the same local perturbation, but only Self-rollout feeds previously corrupted frames back into the model. We evaluate five clean input images with identical actions, seeds, and per-chunk noise in both conditions.

Results. We report LPIPS against the matched clean rollout, averaged over frames within each chunk and across five images. Higher values indicate greater visual deviation.

ChunkSelf rolloutHistory resetΔ\Delta
10.2220.2220.000
20.4350.1950.240
30.6120.2340.378
40.7110.3020.409
  • Self rollout: Local instability + corruption feedback.
  • History reset: Local instability only.
  • Δ\Delta: Additional feedback effect.

Analysis. As we see in the table, with history reset, errors remain relatively small, increasing from 0.222 to 0.302. In contrast, Self-rollout errors grow steadily from 0.222 to 0.711. The gap between the two settings also widens across chunks, reaching 0.409 LPIPS in the final chunk.

These results show that autoregressive feedback amplifies errors beyond local denoising instability. Small errors introduced in early chunks can accumulate through subsequent rollouts, leading to much larger deviations over time.

>>> W4&Q3 Control-Adherence Evaluation

>>> A4 Thanks for this practical suggestion! We add two camera-control adherence metrics to directly evaluate whether generated rollouts follow the commanded actions.

Metrics.

  • RotErr: Deviation of the generated camera rotation from the commanded rotation, measured in degrees.
  • TransErr: Deviation of the generated camera movement from the commanded path, normalized by the commanded path length.
InputRotErr (°) ↑TransErr ↑
Clean16.820.389
BadWorld (Velocity-Min)92.750.573

(Higher values indicate worse control adherence.)

These results show that BadWorld severely disrupts adherence to commanded actions. Together with the visual and geometric metrics reported in the paper, they provide direct evidence that the attack compromises both visual predictions and world-model dynamics.

>>> W5 Robustness and Matched-Compute Evaluation

>>> A5 Thanks for this valuable suggestion! The goal of bi-level optimization is to maintain effective attacks under unknown and challenging future controls, rather than to improve every metric substantially. We examine this benefit from two perspectives.

1. Matched-compute comparison. We evaluate another 10 images under five held-out control trajectories. Since CMA-ES performs additional trajectory evaluations, bi-level with 300 steps and Velocity-Min with 500 steps require comparable attack time. Under matched compute, Bi-level performs better on all six metrics.

MethodI2V-BG ↓BGC ↓Aesth. ↓Imaging ↓CLIP-I ↓MEt3R ↑
Bi-level, 300 steps0.9310.8260.3420.5390.6810.198
Velocity-Min, 500 steps0.9330.8310.3920.5460.6900.180

2. Worst-case robustness. We further compare the metric-wise worst cases over the evaluated image–control pairs. For ↓ metrics, we report the largest value; for MEt3R ↑, we report the smallest value. Bi-level also achieves a better worst-case value on all six metrics.

MethodI2V-BG ↓BGC ↓Aesth. ↓Imaging ↓CLIP-I ↓MEt3R ↑
Bi-level, 300 steps0.9410.8850.3910.6150.8040.133
Velocity-Min, 500 steps0.9470.8930.4400.6540.8230.143

Under matched compute, Bi-level produces stronger average degradation; on the worst-case pairs, it also remains stronger across all metrics. This directly demonstrates that CMA-ES trajectory mining improves both attack strength and robustness to difficult controls.

AuthorsOpenReview ↗

Response to Reviewer MDJc [3/4]

>>> W6&Q5 Driving Safety Claims and Task-Level Evaluation

>>> A6 Valuable suggestions! We agree that task-level evaluation is important for assessing driving safety beyond visual degradation. As suggested, we have expanded the Epona evaluation from 20 to 50 clips spanning 50 different logs, re-evaluated the paper’s metrics, and added three task-related metrics.

Setting
We retain the paper’s attack configuration and use matched clean/adversarial rollouts. The added metrics follow a 10-second recorded-agent evaluation.

Added Metrics

  • Reference-path Progress (RP): How far the vehicle progresses along the ground-truth route, determined by projecting its final position onto that route.
  • Collision Rate (CR): Percentage of clips in which the ego vehicle’s bounding box overlaps with another object’s bounding box at any time.
  • Drivable-area Violation rate (DAV): Percentage of clips with any vehicle corner at least 0.3 m outside the drivable area.

Results

Table 1 here reports video/trajectory errors and the added task-level metrics. Table 2 compares clean and adversarial outputs.

Comparison with Ground Truth and Task-Level Evaluation (Table 1)

InputFVD₁₀ ↑FID ↑Avg.L2 (1/2/3s) ↑†RP@10s ↓†CR≤10s ↑†DAV≤10s ↑
Clean621.4037.251.3142.1210%8%
Adversarial3705.00363.653.4119.5240%16%

(† denotes newly added metrics.)

Direct Comparison Between Clean and Adversarial Outputs (Table 2)

ComparisonOutput LPIPS ↑Initial-plan ΔADE ↑Rollout separation@10s ↑
Clean vs. Adversarial0.6520.980 m26.01 m

Beyond visual degradation and trajectory prediction errors, BadWorld increases 10-second collision rate from 10% to 40%, reduces reference-path progress, and doubles drivable-area violations. These additional task-level results demonstrate severe safety-related consequences under the evaluated setting.


Additional Weaknesses

>>> W1 Statistical Reliability

>>> A1 Thanks! As suggested, we re-evaluate Table 5 with three random seeds and report the mean ± 95% CI below. Bi-level achieves better mean performance on all six metrics, with comparable confidence intervals across the two methods.

MethodI2V-BG ↓BGC ↓Aesth. ↓Imaging ↓CLIP-I ↓MEt3R ↑
Velocity-Min0.937 ± 0.0120.847 ± 0.0100.442 ± 0.0280.558 ± 0.0230.723 ± 0.0150.251 ± 0.015
Bi-level0.930 ± 0.0100.835 ± 0.0160.414 ± 0.0200.556 ± 0.0150.705 ± 0.0100.260 ± 0.013

It is also worth noting that our evaluation already covers diverse controls: each image in Table 1 is tested with 3 distinct camera trajectories, while each of the 10 hard samples in Table 5 is tested with 7 trajectories. Together with the repeated-seed evaluation above, these results show that the bi-level advantage is stable across both random seeds and future controls.

>>> W2 Quantifying Perturbation Imperceptibility

>>> A2 Thanks for this practical suggestion. We provide quantitative measurements and clarify the pixel-range convention as follows:

  • Quantitative evaluation: We report PSNR, SSIM, and LPIPS between clean and adversarial input images for attacks on Astra. The table reports mean ± standard deviation.
MethodPSNR ↑SSIM ↑LPIPS ↓
BadWorld34.981 ± 0.3350.9037 ± 0.03080.1352 ± 0.0566

(Arrows indicate better preservation of the original image.)

  • Pixel-range convention: Our budget ϵ=0.05\epsilon=0.05 is applied in the normalized [−1,1][-1,1] pixel range, equivalent to 0.0250.025 in [0,1][0,1], or approximately 6.375/2556.375/255.

These results quantify the perturbation's visual impact and show that the attacked inputs retain high image fidelity.

Comments on the paper

2 comments

Dear Reviewers,

Thank you so much for your time and constructive feedback. With the discussion deadline coming up, we’d appreciate any thoughts on our rebuttal and additional experiments. We’d be happy to answer any remaining questions and would welcome suggestions for revising the paper.

Thanks again for reviewing our work!

Best regards,

The Authors

AuthorsOpenReview ↗

Summary of Discussions

Dear Area Chair,

Thank you for your time and care in handling our paper. We also sincerely thank all three reviewers for their constructive feedback. As the discussion period draws to a close, we would like to summarize the main points below.


Paper Highlights

BadWorld targets interactive world models to forge small, control-agnostic perturbations on the initial image, thereby disrupting the entire generated video. It combines a self-supervised attack on early denoising velocity with trajectory-adaptive bi-level optimization, ensuring the attack generalizes across unpredictable future controls without needing future ground truth.

As discussed in the paper and rebuttal, this work has two practical motivations. First, representative interactive world models are fragile. Tiny perturbations can break both their visual predictions and physical dynamics. Conversely, these subtle perturbations can also protect personal or copyrighted images from unauthorized video generation. Second, safety risks for real-world decision making are realistic. Exposing these flaws early is critical for safe deployment in settings like autonomous driving. Attackers can either inject digital perturbations into transmitted camera streams or place printed patches directly in the physical scene to mislead these models.


Strengths highlighted by the reviewers

  • Timely and important problem setting (MDJc, DSmP).
  • Well-designed and effective method (MDJc, NYR3).
  • Clear and accessible writing (DSmP).
  • Broad evaluation settings (MDJc)

Main additions and clarifications during the discussion

  • Additional analyses and experiments. We compared BadWorld with I2VGuard and Immune2V, tested bi-level optimization on held-out controls under matched compute, and expanded the cross-model transfer evaluation. We also measured perturbation fidelity and detectability, evaluated the surrogate using generated histories, and isolated the effect of autoregressive feedback through a history-reset experiment.
  • Localized-patch extension. For autonomous driving, we extended our digital-image attack to simulate a patch placed on the road within the camera’s view. On Epona, this image-space patch raised the measured 10-second collision rate from 10% to 30%. We also provide visuals to directly justify this: https://anonymouspage123.github.io/badworld-page/

Discussion outcome

Reviewers MDJc and NYR3 each assigned a score of 2. We responded to both reviews and continued the discussion with NYR3. Reviewer DSmP indicated that our response addressed the concerns and maintained score of 3.

We have posted detailed responses in the individual review threads and incorporated the changes into the revised manuscript. All revisions are highlighted in blue for ease of reference.


Thank you again for the time and care you have devoted to our paper. We sincerely appreciate your consideration and the opportunity to discuss our work with the reviewers.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.