Research map / Adversarial robustness and security / Attacks on text and graphs
Jailbreaking Multimodal Large Language Models via Shuffle Inconsistency
Shiji Zhao, Ranjie Duan, Fengxiang Wang, Chi Chen, Caixin Kang, Shouwei Ruan, Jialing Tao, YueFeng Chen, Hui Xue, Xingxing Wei
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 |
|---|---|---|---|
| shuffle harmful text words | original harmful text prompts · MM-SafetyBench 01-IA, GPT-4o | toxic score: 1.51 -> 3.19 | Improved |
| shuffle harmful image patches | original harmful images · MM-SafetyBench 01-IA, GPT-4o | toxic score: 1.51 -> 2.13 | Improved |
| shuffle image and text, vs other mutations | text mutation · MM-SafetyBench 01-IA | toxic score: 2.75 -> 3.19 | Improved |
| add image shuffling (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | ASR (%): 13.40 -> 35.05 | Improved |
| add text shuffling (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | ASR (%): 13.40 -> 67.01 | Improved |
| add image shuffling to text shuffling (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | ASR (%): 67.01 -> 80.41 | Improved |
| add query-based optimization to random shuffling (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | ASR (%): 28.87 -> 80.41 | Improved |
| add query-based optimization to random shuffling (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | toxic score: 2.65 -> 3.96 | Improved |
| increase max query iteration from 1 to 10 (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | ASR (%): 28.87 -> 80.41 | Improved |
| increase max query iteration from 10 to 20 (ablation) | SI-Attack · MM-SafetyBench 01-IA no typography, GPT-4o | ASR (%): 80.41 -> 81.44 | No clear effect |
| use SI-Attack instead of QR-Attack | jailbreak attack · MM-SafetyBench on GPT-4o | ASR (%): 20.77 -> 68.57 | Improved |
| use SI-Attack instead of QR-Attack | jailbreak attack · MM-SafetyBench on VLGuard | ASR (%): 9.52 -> 40.77 | Improved |
Closest papers by meaning
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.92)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.92)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.91)
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.90)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.90)
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.
