Structured-based Jailbreak Attack Defense on MM-SafetyBench unseen attack types
2.35ASR (SD)NullSteer
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| NullSteerBackbone=MiniGPT-42026.03 | 2.35 | 6.28 | 11.57 | |
| ASTRABackbone=MiniGPT-42026.03 | 3.75 | 8.75 | 11.25 | |
| NullSteerBackbone=LLaVA-v1.52026.03 | 7.65 | 24 | 5.8 | |
| ASTRABackbone=LLaVA-v1.52026.03 | 8.75 | 25 | 6.25 | |
| NullSteerBackbone=Qwen2-VL2026.03 | 9.86 | 38.5 | 32.53 | |
| ASTRABackbone=Qwen2-VL2026.03 | 11.25 | 40 | 33.75 | |
| w/o defenseBackbone=MiniGPT-42026.03 | 13.75 | 43.25 | 43.75 | |
| w/o defenseBackbone=LLaVA-v1.52026.03 | 18.75 | 55 | 22.5 | |
| w/o defenseBackbone=Qwen2-VL2026.03 | 20 | 61.25 | 38.75 |