Research map / Adversarial robustness and security / Attacks on text and graphs
From ``Sure" to ``Sorry": Detecting Jailbreak in Large Vision Language Model via JailNeurons
Yuyou Gan, Qingming Li, Junhao Li, Zhi Chen, Jinbao Li, Xiaoming Li, Shouling Ji
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 JailNeuron mask filtering (vs no filtering) (ablation) | JDJN on LLaVA · Normal benign | FPR: 0.442 -> 0.019 | Improved |
| use JailNeurons instead of PCA (ablation) | JDJN on LLaVA · Normal benign | FPR: 0.626 -> 0.019 | Improved |
| use JailNeurons instead of SNIP selection (ablation) | JDJN on LLaVA · Normal benign | FPR: 0.577 -> 0.019 | Improved |
| use JailNeurons instead of NFNR (ablation) | JDJN on LLaVA · JAMLLM | TPR: 0.874 -> 0.942 | Improved |
| use linear SVM instead of MLP detector (ablation) | JDJN on Janus-pro · FigStep | TPR: 0.781 -> 1.0 | Improved |
| use linear SVM instead of non-linear SVM (ablation) | JDJN on Janus-pro · JAMLLM | TPR: 0.783 -> 0.853 | Improved |
| use refusal token sorry instead of unfortunately (ablation) | JDJN1 on Janus-pro · Normal benign | accuracy: 0.722 -> 0.956 | Improved |
| apply mask threshold tau>0 instead of no mask (ablation) | JDJN1 · six distributions | worst-case accuracy | Improved |
| raise lambda from 0.1 to 0.5 (ablation) | JDJN on LLaVA · Normal benign | accuracy: 0.94+ -> 0.73 | Hurt |
| train on JailBreakV instead of FigStep (JDJN1 vs JDJN2) | JDJN · JAMLLM on LLaVA | TPR: 0.524 -> 0.942 | Improved |
| train benign set on MM-Vet instead of Normal | JDJN · Normal benign on LLaVA | FPR: 0.346 -> 0.019 | Improved |
| train benign set on MM-Vet instead of MM-Bench | JDJN · MM-Vet on LLaVA | FPR: 0.285 -> 0.0 | Improved |
Closest papers by meaning
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.91)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.89)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.89)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.89)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 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.
