Research map / Adversarial robustness and security / Attacks on text and graphs
Heuristic-Induced Multimodal Risk Distribution Jailbreak Attack for Multimodal Large Language Models
Teng Ma, Xiaojun Jia, Ranjie Duan, Xinfeng Li, Yihao Huang, Xiaoshuang Jia, Zhixuan Chu, Wenqi Ren
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 heuristic-induced search for understanding-enhancing prompt p_u (ablation) | HIMRD · SafeBench (LLaVA-V1.5) | attack success rate (%): 62.57 -> 73.71 (+11.14) | Improved |
| add heuristic-induced search for inducing prompt p_i after p_u (ablation) | HIMRD · SafeBench (LLaVA-V1.5) | attack success rate (%): 73.71 -> 82.29 (+3.06) | Improved |
| add heuristic-induced search for p_u (ablation) | HIMRD · SafeBench (GLM-4V-9B) | attack success rate (%): 86.29 -> 89.35 (+3.06) | Improved |
| add heuristic-induced search for p_i after p_u (ablation) | HIMRD · SafeBench (GLM-4V-9B) | attack success rate (%): 89.35 -> 94.29 (+4.94) | Improved |
| increase search iterations N1 for p_u from 1 to 2 (ablation) | HIMRD · SafeBench (Qwen-VL-Chat) | attack success rate (%): roughly 12% | Improved |
| increase search iterations N2 for p_i up to five (ablation) | HIMRD · SafeBench (GLM-4V-9B) | attack success rate (%): approximately 5% | Improved |
| distribute malicious prompt across text and image instead of single modality (ablation) | HIMRD · SafeBench (GLM-4V-9B refusal) | refusal rate (%): 76.00 -> 0.00 (text), 30.00 -> 0.2 (vision) | Improved |
| apply image denoising and perplexity-based text defense | HIMRD · SORRY-Bench | attack success rate (%): 95.56 -> 95.11 (Qwen-VL-Chat), 94.89 -> 95.33 (GLM-4V-9B) | No clear effect |
| use HIMRD instead of FigStep on open-source MLLMs | HIMRD · SafeBench average | attack success rate (%): 63.59 -> 90.86 | Improved |
| use HIMRD instead of FigStep on closed-source MLLMs | HIMRD · tiny-SafeBench average | attack success rate (%): 38.57 -> 68.09 | Improved |
| use HIMRD instead of white-box UMK on MiniGPT-4 | HIMRD · SafeBench (MiniGPT-4) | attack success rate (%): 87.71 -> 84.57 | Hurt |
Closest papers by meaning
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.89)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.89)
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.89)
- h4rm3l: A Language for Composable Jailbreak Attack Synthesis (ICLR 2025 · similarity 0.89)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · 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.
