Research map / Video / Video generation
BadVideo: Stealthy Backdoor Attack against Text-to-Video Generation
Ruotong Wang, Mingli Zhu, Jiarong Ou, Rui Chen, Xin Tao, Pengfei Wan, Baoyuan Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add time-related instruction to the MLLM detector prompt | GPT-4o video detection of BadVideo · LaVie backdoored videos | detection success rate: 12% -> 52% | Improved |
| add time-related instruction to the MLLM detector prompt | Qwen2.5VL-7B detection · LaVie backdoored videos | detection success rate: 0% -> 12% | Improved |
| add time-related instruction to safety-focused moderators | Omni-Moderation and Llama-Guard-3-11B-Vision · LaVie backdoored videos | detection success rate: 0% -> 0% | No clear effect |
| increase number of frames analysed from 1 to 16 (ablation) | GPT-4o detector · LaVie backdoored videos | detection success rate: 2% -> 52% | Improved |
| increase number of frames analysed from 1 to 16 (ablation) | GPT-4o detector · LaVie backdoored videos | tokens and time per sample: 360 -> 4214 tokens; 2.8 -> 12.6 s | Hurt |
| add backdoor fine-tuning on clean data (10% of training set, up to 100 epochs) as defense | BadVideo backdoored model · LaVie | ASR_MLLM: above 80% | No clear effect |
| increase training epochs (ablation) | BadVideo on LaVie · LaVie | ASR_MLLM: over 70% after 80 epochs | Improved |
| increase poisoning ratio from 5% to 30% (ablation) | BadVideo on LaVie · LaVie | ASR_MLLM: above 80% at 20% poisoning ratio | Improved |
| increase prompt perturbation strength up to 80% | BadVideo · LaVie | CPR | Hurt |
| embed three backdoors with different triggers in one model | BadVideo · LaVie | ASR_MLLM | No clear effect |
| use Cyrillic substitution trigger instead of rare word (ablation) | BadVideo STC · LaVie | ASR_Human: 92.3 -> 90.2 | No clear effect |
| use benign phrase trigger instead of rare word (ablation) | BadVideo STC · LaVie | ASR_Human: 92.3 -> 91.7 | No clear effect |
Closest papers by meaning
- T2VSafetyBench: Evaluating the Safety of Text-to-Video Generative Models (NeurIPS 2024 · similarity 0.88)
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks (NeurIPS 2025 · similarity 0.88)
- I2VGuard: Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models (CVPR 2025 · similarity 0.87)
- VMDT: Decoding the Trustworthiness of Video Foundation Models (NeurIPS 2025 · similarity 0.85)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.85)
- AdvI2I: Adversarial Image Attack on Image-to-Image Diffusion Models (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
