Visual Jailbreak Defense on Visual Adversarial Attacks epsilon = 16/255
12.45Attack Success RateIMMUNE
Evaluation Results
| Method | Links | |
|---|---|---|
| IMMUNEModel=LLaVA-1.62024.11 | 12.45 | |
| IMMUNEModel=Qwen-VL2024.11 | 13.22 | |
| IMMUNEModel=LLaVA-1.52024.11 | 13.32 | |
| AdaShieldModel=Qwen-VL2024.11 | 17.32 | |
| IMMUNEModel=MiniGPT-4-7B2024.11 | 18.42 | |
| IMMUNEModel=MiniGPT-4-13B2024.11 | 23.41 | |
| AdaShieldModel=LLaVA-1.62024.11 | 27 | |
| FigStepModel=Qwen-VL2024.11 | 34.25 | |
| AdaShieldModel=LLaVA-1.52024.11 | 35.21 | |
| OriginalModel=Qwen-VL2024.11 | 38.11 | |
| CoCAModel=Qwen-VL2024.11 | 38.15 | |
| AdaShieldModel=MiniGPT-4-7B2024.11 | 39 | |
| CoCAModel=LLaVA-1.62024.11 | 44.23 | |
| CoCAModel=LLaVA-1.52024.11 | 45.21 | |
| FigStepModel=MiniGPT-4-7B2024.11 | 47.09 | |
| AdaShieldModel=MiniGPT-4-13B2024.11 | 47.82 | |
| FigStepModel=LLaVA-1.52024.11 | 49.1 | |
| CoCAModel=MiniGPT-4-7B2024.11 | 53.19 | |
| OriginalModel=MiniGPT-4-7B2024.11 | 53.4 | |
| FigStepModel=LLaVA-1.62024.11 | 53.98 | |
| FigStepModel=MiniGPT-4-13B2024.11 | 59.81 | |
| OriginalModel=MiniGPT-4-13B2024.11 | 62.57 | |
| OriginalModel=LLaVA-1.52024.11 | 66.12 | |
| OriginalModel=LLaVA-1.62024.11 | 68 | |
| CoCAModel=MiniGPT-4-13B2024.11 | 71.85 |