Safety Evaluation on MIS (Hard)
7.33Attack Success Rate (ASR)VLGuard
Evaluation Results
| Method | Links | |
|---|---|---|
| VLGuardBackbone=Qwen2-VL2025.08 | 7.33 | |
| PRISMBackbone=Qwen2-VL2025.08 | 11.29 | |
| PRISMBackbone=LLaVA-1.52025.08 | 20 | |
| SPA-VLBackbone=LLaVA-1.52025.08 | 23.56 | |
| VLGuardBackbone=LLaVA-1.52025.08 | 28.12 | |
| SPA-VLBackbone=Qwen2-VL2025.08 | 30.06 | |
| SafeRLHF-VBackbone=Qwen2-VL2025.08 | 70.3 | |
| SafeRLHF-VBackbone=LLaVA-1.52025.08 | 72.48 | |
| No DefenseBackbone=Qwen2-VL2025.08 | 78.02 | |
| No DefenseBackbone=LLaVA-1.52025.08 | 78.81 |