Safety Defense Evaluation on Jailbreak Attacks (GCG, AutoDAN, PAIR, DeepInc., SAP30, SIJ)
1GCG Attack ScorePPL
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| PPLBackbone=Vicuna-13B2025.10 | 1 | 4.32 | 3.1 | 3.1 | 3.92 | 2.76 | 3.03 | |
| Self-ExaminationBackbone=Vicuna-13B2025.10 | 1 | 3.3 | 1.96 | 2.68 | 1.7 | 1.86 | 2.08 | |
| SAIDBackbone=Vicuna-13B2025.10 | 1 | 1 | 1.16 | 1 | 1.03 | 1 | 1.03 | |
| No DefenseBackbone=Vicuna-13B2025.10 | 1.14 | 4.3 | 3.2 | 3.14 | 3.95 | 2.58 | 3.05 | |
| ParaphraseBackbone=Vicuna-13B2025.10 | 1.14 | 2.96 | 1.96 | 3.46 | 2.7 | 1.56 | 2.3 | |
| ICDBackbone=Vicuna-13B2025.10 | 1.14 | 4.04 | 2.48 | 3.24 | 2.55 | 3.16 | 2.77 | |
| SafeDecodingBackbone=Vicuna-13B2025.10 | 1.24 | 1.32 | 1.8 | 1.2 | 1.74 | 3.62 | 1.82 | |
| RetokenizationBackbone=Vicuna-13B2025.10 | 1.3 | 3.6 | 3.82 | 2.38 | 4.46 | 3.16 | 3.12 |