Adversarial Attack on JailBreakBench (Llama-2-7B non-trivial subset)
26ASRPGD-10
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PGD-10Gradient Type=Exact gradient2026.05 | 26 | 0.9 | 0.23 | |
| PGD-5Gradient Type=Exact gradient2026.05 | 24 | 1.7 | 0.42 | |
| PGD-2Gradient Type=Exact gradient2026.05 | 14 | 4.3 | 0.6 | |
| RS-FGSMGradient Type=Exact gradient2026.05 | 12 | 8.5 | 1.03 | |
| FGSMGradient Type=Exact gradient2026.05 | 6 | 8.8 | 0.53 | |
| FGMGradient Type=Exact gradient2026.05 | 0 | 8.9 | 0 | |
| FGSMGradient Type=GPR (predicted)2026.05 | 0 | 70.2 | 0 | |
| FGMGradient Type=GPR (predicted)2026.05 | 0 | 46.1 | 0 | |
| RS-FGSMGradient Type=GPR (predicted)2026.05 | 0 | 70 | 0 | |
| PGD-2Gradient Type=GPR (predicted)2026.05 | 0 | 37.6 | 0 | |
| PGD-5Gradient Type=GPR (predicted)2026.05 | 0 | 15.7 | 0 | |
| PGD-10Gradient Type=GPR (predicted)2026.05 | 0 | 8 | 0 |