Research map / Adversarial robustness and security / Attacks on text and graphs
Jailbreaking on Text-to-Video Models via Scene Splitting Strategy
Wonjun Lee, Haon Park, Doehyeon Lee, Bumsub Ham, Suhyun Kim
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 Scene Manipulation (iterative single-scene modification) to Scene Splitting (ablation) | SceneSplit · T2VSafetyBench, Veo2 | ASR: 42.7% -> 60.9% | Improved |
| add Strategy Update (outer loop plus Strategy Library) (ablation) | SceneSplit · T2VSafetyBench, Veo2 | ASR: 60.9% -> 78.2% | Improved |
| add Strategy Library with iterations held fixed (ablation) | SceneSplit · T2VSafetyBench, Veo2 | ASR: 69.1% -> 78.2% | Improved |
| split prompt into multiple scenes (Scene Division) alone (ablation) | T2VSafetyBench prompts · one-shot ASR, Veo2 and Hailuo | ASR: Veo2 33.1% -> 37.7%; Hailuo 40.9% -> 53.6% | Improved |
| apply only paraphrasing to soften language, without scene division (ablation) | T2VSafetyBench prompts · one-shot ASR, Veo2 and Hailuo | ASR: Veo2 33.1% -> 30.0%; Hailuo 40.9% -> 42.7% | Mixed |
| combine scene division with paraphrasing (ablation) | Scene Division · one-shot ASR, Veo2 and Hailuo | ASR: Veo2 37.7% -> 42.7%; Hailuo 53.6% -> 56.4% | Improved |
| use Scene Manipulation instead of regenerating new scene splits each iteration (ablation) | SceneSplit · Veo2, 5 iterations, no Strategy Update | ASR: 46.4% -> 60.9% | Improved |
| increase outer iterations T from 1 to 3 (ablation) | SceneSplit · Veo2 | ASR: 60.9% -> 73.6% -> 78.2% | Improved |
| increase inner iterations I from 3 to 8 (ablation) | SceneSplit · Veo2 | ASR: 57.7% -> 60.9% -> 62.4% | Improved |
| add Strategy Update for efficiency (ablation) | SceneSplit · Veo2 | average attack attempts: 6.22 -> 5.54 attempts, time 311.52s -> 323.06s | Improved |
| raise strategy retrieval similarity threshold from 0.6 to 0.8 (ablation) | SceneSplit · Veo2 | ASR: 78.2% -> 72.7% | Hurt |
| use non-fixed initial scene split instead of fixed (ablation) | SceneSplit · Veo2 | ASR: 78.2% -> 77.3% | No clear effect |
Closest papers by meaning
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks (NeurIPS 2025 · similarity 0.91)
- T2VSafetyBench: Evaluating the Safety of Text-to-Video Generative Models (NeurIPS 2024 · similarity 0.91)
- ColJailBreak: Collaborative Generation and Editing for Jailbreaking Text-to-Image Deep Generation (NeurIPS 2024 · similarity 0.87)
- Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs (NeurIPS 2025 · similarity 0.87)
- ART: Automatic Red-teaming for Text-to-Image Models to Protect Benign Users (NeurIPS 2024 · similarity 0.86)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.86)
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.
