Jailbreak Attack on SafetyBench LLaVA-2 Integrated from AdvBench (test)
83.73Illegal Activity Success RateMLAI
Evaluation Results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MLAIAttack Type=White-box2024.11 | 83.73 | 80.36 | 79.1 | 79.62 | 50.18 | 75.85 | 68.98 | 93.45 | 74.82 | 93.59 | 96.77 | 100 | 100 | 82.8 | |
| Liu et al.Attack Type=White-box2024.11 | 74.77 | 40.97 | 64.63 | 61.04 | 18.42 | 57.2 | 61.23 | 91.15 | 49.36 | 95.46 | 97.21 | 100 | 98.97 | 70.03 | |
| Qi et al.Attack Type=White-box2024.11 | 48.03 | 40.21 | 53.85 | 50.31 | 7.98 | 47.97 | 41.55 | 85.62 | 37.33 | 82.94 | 87.69 | 90.04 | 91.22 | 58.83 | |
| Plain TextAttack Type=White-box2024.11 | 3.82 | 3.08 | 30.75 | 13.88 | 4.15 | 4.97 | 20.64 | 74.59 | 20.92 | 80.16 | 87.48 | 85.61 | 87.74 | 39.83 |