acceptodds
Withdrawn from ICLR 2027

BadAction: Backdoor Attacks on Interactive Video Generation via Action-Guided Triggers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Interactive video generation (IVG) models have achieved remarkable progress in producing controllable visual content guided by user-defined actions, yet their security vulnerabilities remain largely unexplored. In this paper, we present the first systematic study of backdoor attacks against the interactivity of IVG models. Based on this attack surface, we propose BadAction, which leverages action-guided triggers to achieve the attack. Specifically, BadAction implants predefined motion patterns into the action sequences of backdoor samples and associates them with a static target video. Once triggered, the backdoored model generates frozen future frames that no longer respond to subsequent user actions, while preserving normal behavior on benign action sequences. In addition, we also explore a stealthier attack where the multimodal triggers by jointly poisoning multiple modalities. Experiments show that BadAction achieves average attack success rates of 91.0% with action-only triggers and 80.4% with multimodal triggers. Moreover, extensive defense evaluations show that BadAction successfully bypasses existing backdoor detection methods, revealing a critical security gap in the interactive video generation pipeline.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
interactive video generation, backdoor attack, action control, multimodal trigger, diffusion models, AI security

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 4 reviews, above 27% of the others.
Reviewer_pNt9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces BadAction, a backdoor attack framework for interactive video models. Unlike prior attacks on image-generation diffusion models, BadAction uses interactive actions, typically camera movements, as triggers. A specific interaction command activates a hidden backdoor that freezes subsequent video frames. Experiments demonstrate support for both single-interaction and multimodal triggers, with an average attack success rate exceeding 80%. The work offers a promising exploration of backdoor vulnerabilities in interactive video generation.

Critical strengths

Strength. 1 Strong Motivation

While prior work has extensively studied backdoor attacks on diffusion models for image generation, backdoor threats in downstream applications remain relatively underexplored. The authors address this gap by proposing an effective backdoor attack framework for diffusion-based interactive video models.

Strength. 2 Exploration of Multimodal Triggers

Prior backdoor attacks typically rely on triggers within a single modality, such as text or image patches. The authors extend BadAction to incorporate multimodal triggers and examine how this design affects both attack effectiveness and defense performance.

Critical weaknesses

Weakness. 1 Substantial Degradation in Generation Quality

The reported FVD increases from 345.7 for the baseline model to 759.1 for BadAction, indicating substantial degradation in video generation quality during normal operation. The multimodal variant exhibits an even larger increase in FVD, exceeding five times the baseline value. These results call into question the authors’ claim that the attack preserves the model’s normal functionality.

Weakness. 2 Limited Stealth and Incomplete Defense Evaluation

The defense experiments show that T2IShield and UFID, originally designed for image generation tasks, can still detect a substantial proportion of backdoor samples. These results raise concerns about BadAction’s ability to evade existing defenses. Furthermore, the evaluation omits more recent backdoor detection methods, such as STEDiff [1] and TERD [2]. Including these methods, with appropriate adaptations where necessary, would provide a more comprehensive assessment of the attack’s stealth.

Weakness. 3 Single Baseline Model

The authors only evaluated BadAction within a single framework, which does not provide evidence for the generalization ability of the attack algorithm.

Questions for authors

Please refer to Weaknesses.

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

This paper investigates backdoor attacks on interactive video generation (IVG) models, focusing on the action control channel rather than conventional text or image triggers. The proposed method, BadAction, replaces a subsequence of actions in poisoned training samples with a predefined four-step camera-motion trigger and pairs it with a target video that becomes static after trigger activation. The model is fine-tuned using a combined benign and poisoned loss, aiming to make the generated video ignore subsequent user actions once triggered while preserving normal behavior on benign inputs. The approach is further extended to multimodal triggers involving action, text, and image inputs. Experiments are conducted on Astra (based on Wan2.1-T2V-1.3B) using a subset of the Sekai dataset. Attack effectiveness is evaluated using SSIM-based and human-judged attack success rates (ASR), while benign generation quality is assessed using FVD and CLIPSIM. BadAction achieves a 91.0% SSIM-based ASR with action-only triggers, outperforming image- and text-based baselines. The paper also presents ablation studies and evaluates resistance to existing backdoor detection methods, including T2IShield and UFID.

Critical strengths

The paper explores a previously underexamined backdoor attack surface in interactive video generation (IVG) models by introducing action-sequence triggers that disrupt interactivity, causing generated videos to freeze and ignore subsequent user actions.

Critical weaknesses
  1. In Table 2, the FVD on benign inputs increases from 345.7 for the benign model to 759.1 for BadAction, representing approximately a 120% increase. Such a substantial degradation raises concerns about whether the model utility objective stated in Section 3.2 is adequately achieved. The authors argue that generation utility is preserved because BadAction achieves FVD scores comparable to other backdoor baselines. However, utility preservation should also be assessed relative to a model without backdoor poisoning. Furthermore, the source of this degradation remains unclear. The paper does not specify whether the benign model is the original Astra checkpoint or a model fine-tuned without poisoning under identical conditions. Notably, BadNets-T, which uses only a text trigger, exhibits a nearly identical FVD of 758.1. This suggests that the degradation may stem from factors shared across the fine-tuning procedures, such as the limited training dataset (100 clips) or the use of static target videos in poisoned samples, rather than the action trigger itself. A clean fine-tuning baseline under identical training conditions is needed to isolate the impact of poisoning on generation quality.

  2. FVD and CLIPSIM measure distribution-level video quality and video–text alignment, respectively, but neither directly evaluates whether generated videos follow the input actions. Since action controllability is precisely the functionality targeted by BadAction, preserving this capability on benign inputs is essential to the claimed model utility. However, the paper provides no quantitative evaluation of action-following performance on benign inputs, leaving the claimed preservation of normal interactivity insufficiently supported. Comparing action-following performance before and after backdoor fine-tuning would substantially strengthen this claim. For example, the authors could measure the discrepancy between estimated camera trajectories in generated videos and the input action trajectories, or conduct human evaluations of action-following accuracy.

Questions for authors

Q1. In Table 2, FVD increases from 345.7 to 759.1 (approximately 120%) on benign inputs. Prior video backdoor research, such as BadVideo [1], reports relatively small FVD changes compared to corresponding non-backdoored models. What accounts for this substantial degradation? Please clarify whether the benign model is the original pre-trained Astra model and report the FVD of a clean fine-tuned model under identical training conditions without poisoning.

Q2. FVD and CLIPSIM in Table 2 do not directly measure whether generated videos follow the input actions. Could the authors provide a quantitative comparison of action-following performance on benign inputs before and after backdoor fine-tuning, such as camera trajectory errors?

Q3. As noted in Section 2.2, T2IShield and UFID primarily target text- or image-based backdoors rather than the action control channel. How were these defenses adapted to IVG, and could their low detection rates in Section 4.4 reflect this modality mismatch rather than the inherent stealthiness of BadAction? In particular, which input modalities were perturbed in UFID? Additionally, the reported results appear inconsistent with the stated fixed 5% FPR protocol. For example, the UFID tri-modal result in Table 4 (Precision 30.0%, Recall 15.0%) implies a 35% FPR if computed from a single evaluation set containing 100 positive and 100 negative samples. Some recalls in Table 3 also have fractional values that cannot arise from a single evaluation of 100 samples. Could the authors clarify the threshold selection and metric aggregation procedures?

Q4. The paper claims that composite triggers activate only when all specified modality triggers are present (Sections 3.3.3 and 4.2.2). However, Table 1 alone does not establish whether each modality is necessary for activation. Since the ASRs of action-containing combinations (80.4–84.7%) are closer to the action-only result (91.0%) than to the image-text result (58.5%), could the attack be predominantly driven by the action trigger? Could the authors report ASRs when only the action trigger, or other partial trigger combinations, are provided to the same composite-backdoored model?

Q5. In Figure 3, the benign sequences appear to contain the same W → A → W → D key pattern as the four-step trigger defined in Section 4.1, yet they do not produce frozen videos. Were the benign and triggered sequences generated using the same backdoored checkpoint? If so, how do the benign actions differ from the trigger in terms of the actual camera displacement values?

[1] BadVideo: Stealthy Backdoor Attack against Text-to-Video Generation. ICCV 2025.

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

This paper introduces BadAction, a backdoor attack method for IVG models. By poisoning training data with a predefined action pattern, the attack causes the model to generate frozen frames once triggered.

Critical strengths
  1. Investigating security risks in the action-control IVG models sounds meaningful.

  2. The proposed method is clear.

  3. The writing is good and the structure is easy to follow.

  4. The experiments look abundant.

Critical weaknesses
  1. Technical novelty should be strengthened. From my perspective, the attack appears to be a direct extension of standard conditional data poisoning to action inputs. It is unclear what unique technical challenges are introduced by the action channel.

  2. The evidence for preserving benign utility is insufficient. The evaluation is conducted on a single model architecture and one dataset. More importantly, the claim that “Badaction preserves generation utility” is not convincingly supported by the results: Table 2 shows the FVD degrades from 345.7 to 759.1.

  3. The training process is unexplained. In Figure 6(d), the attack success rate peaks at epoch 8 and then drops. What causes the decline?

Questions for authors
  1. Technical novelty should be strengthened. From my perspective, the attack appears to be a direct extension of standard conditional data poisoning to action inputs. It is unclear what unique technical challenges are introduced by the action channel.

  2. The evidence for preserving benign utility is insufficient. The evaluation is conducted on a single model architecture and one dataset. More importantly, the claim that “Badaction preserves generation utility” is not convincingly supported by the results: Table 2 shows the FVD degrades from 345.7 to 759.1.

  3. The training process is unexplained. In Figure 6(d), the attack success rate peaks at epoch 8 and then drops. What causes the decline?

Reviewer_1sqwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BadAction studies backdoor attacks on interactive video generation by pairing predefined action sequences with videos that freeze after triggering. Experiments on Astra evaluate action-only and multimodal triggers, benign generation quality, and resistance to two existing detectors.

Critical strengths
  • The action-control channel is a relevant attack surface, and loss of interactivity provides a clear behavioral objective.
  • The evaluation considers several trigger configurations and complements automatic metrics with human judgments.
Critical weaknesses
  • Utility preservation is insufficiently supported. FVD increases from 345.7 to 759.1 for BadAction and 2050.8 for the tri-modal attack. These substantial increases need explanation, while FVD and CLIPSIM do not directly measure whether benign action-following remains accurate.
  • Numerical motion parameters and activation tolerance are underspecified. Similar or incomplete action sequences are not evaluated, and partial-trigger controls are missing for multimodal models. Consequently, the necessity of the complete trigger is not established.
  • Different configurations update different modules. Without accounting for optimization budgets and benign-utility costs, the results do not establish that the action channel is intrinsically the most vulnerable.
  • Adaptations of T2IShield and UFID to action-conditioned video are underspecified. The stated sample counts and 5% FPR also require reconciliation with T2IShield’s reported 98% precision and 50% recall.
Questions for authors
  • What checkpoint is the benign reference in Table 2? Is its fine-tuning budget matched, and what evidence supports preserved action-following despite the FVD increase?
  • What numerical parameters define the action trigger? Were similar or incomplete sequences, partial multimodal triggers, and the clean model under identical inputs evaluated?
  • How are differences in trainable modules and benign-utility costs accounted for when concluding that the action channel is most vulnerable?
  • How are the detectors adapted to video and action inputs? Please clarify the achieved test FPR and the confusion counts or aggregation procedure underlying Table 3.

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

Related papers

Discussion (0)

Sign in to comment.