Research map / Video / Video generation
T2VShield: Model-Agnostic Jailbreak Defense for Text-to-Video Models
Siyuan Liang, Jiayang Liu, Jiecheng Zhai, Tianmeng Fang, Rongcheng Tu, Aishan Liu, Xiaochun Cao, Dacheng Tao
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 multi-scope output detection to the LLaVA detector (ablation) | LLaVA · T2VSafetyBench jailbreak (Open-Sora) | ASR: 51.1 -> 42.2 | Improved |
| add RiskTrace CoT input rewriting (ablation) | LLaVA · T2VSafetyBench jailbreak (Open-Sora) | ASR: 51.1 -> 44.7 | Improved |
| add PosNegRAG retrieval on top of RiskTrace CoT (ablation) | LLaVA + RiskTraceCoT · T2VSafetyBench jailbreak (Open-Sora) | ASR: 44.7 -> 29.7 | Improved |
| add multi-scope detection to complete T2VShield (ablation) | LLaVA + RiskTraceCoT + PosNegRAG · T2VSafetyBench jailbreak (Open-Sora) | ASR: 29.7 -> 22.4 | Improved |
| add RiskTrace CoT, PosNegRAG and multi-scope detection (T2VShield) | Undefended Open-Sora · T2VSafetyBench jailbreak (Open-Sora) | ASR: 52.9% -> 17.2% | Improved |
| add T2VShield | Undefended CogVideoX · SafeWatch | ASR: 47.8% -> 12.3% | Improved |
| add T2VShield | Undefended Keling · Commercial T2V jailbreak | ASR: 75.5% -> 9.1% | Improved |
| add adversarial training to the T2V model | Open-Sora · Clean prompts | FVD: 639.1 -> 747.1 | Hurt |
| add unlearning defense to the T2V model | CogVideoX · Clean prompts | FVD: 474.6 -> 1469.5 | Hurt |
| add keyword detection input filter | CogVideoX · T2VSafetyBench jailbreak | ASR: 55.6% -> 44.1% | Improved |
| add implicit meaning analysis input filter | CogVideoX · T2VSafetyBench jailbreak | ASR: 55.6% -> 54.5% | No clear effect |
| add output-level visual classification | CogVideoX · T2VSafetyBench jailbreak | ASR: 55.6% -> 54.9% | No clear effect |
Closest papers by meaning
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks (NeurIPS 2025 · similarity 0.85)
- BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks (ICLR 2025 · similarity 0.85)
- Perception-Guided Jailbreak Against Text-to-Image Models (AAAI 2025 · similarity 0.84)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.82)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.82)
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.82)
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.
