Research map / Image generation / Text-to-image diffusion
Reason2Attack: Jailbreaking Text-to-Image Models via LLM Reasoning
Chenyu Zhang, Lanjun Wang, Yiwen Ma, Wenhui Li, Guoqing Jin, Anan Liu
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 SFT on Frame Semantics CoT examples (ablation) | Llama-3-8b-Instruct attacker LLM · SD1.4 + safety filters | average ASR: 0.49 -> 0.74 | Improved |
| add RL with attack result reward (ablation) | Llama-3-8b-Instruct attacker LLM · SD1.4 + safety filters | average ASR: 0.49 -> 0.74 | Improved |
| use attack process reward instead of attack result reward in RL (ablation) | RL-trained attacker LLM · SD1.4 + safety filters | average ASR: 0.74 -> 0.87 | Improved |
| use attack process reward instead of attack result reward in RL (ablation) | RL-trained attacker LLM · SD1.4 + safety filters | average query number: 3.3 -> 2.6 | Improved |
| add SFT on CoT examples before RL with attack process reward (ablation) | LLM+RL_AP · SD1.4 + safety filters | average ASR: 0.87 -> 0.90 | Improved |
| add SFT on CoT examples before RL with attack process reward (ablation) | LLM+RL_AP · SD1.4 + safety filters | average query number: 2.6 -> 2.5 | Improved |
| add SFT on CoT examples before RL with attack process reward (ablation) | LLM+RL_AP · SD1.4 + safety filters | average perplexity: 122 -> 155 | Hurt |
| use attack result reward only in RL (ablation) | RL training of attacker LLM · RL return curve | training stability | Hurt |
| R2A compared with iterative CMMA baseline | R2A · SD1.4 + safety filters | average query number: 21.7 -> 2.5 | Improved |
| smaller gamma-style length/stealth trade-off factor (ablation) | R2A · SD1.4 + safety filters | average ASR | Improved |
| increase maximum number of queries (ablation) | R2A · SD1.4 + safety filters | average ASR | Improved |
| R2A on text-to-video model OpenSora V2 versus CMMA | R2A · OpenSora V2 | average ASR: 0.59 -> 0.62 | Mixed |
Closest papers by meaning
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks (NeurIPS 2025 · similarity 0.91)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.90)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.90)
- Improved Generation of Adversarial Examples Against Safety-aligned LLMs (NeurIPS 2024 · similarity 0.90)
- Adversarial Reasoning at Jailbreaking Time (ICML 2025 · similarity 0.90)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.89)
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.
