Jailbreak Defense on GCG adversarial prompts (test)
98DSRBacktranslation w/o early termination
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Backtranslation w/o early terminationTarget Model=Vicuna-13B, early termination=false2024.02 | 98 | 78.22 | |
| Backtranslation w/ early terminationTarget Model=Vicuna-13B, early termination=true2024.02 | 96 | 61.7 | |
| SmoothLLMTarget Model=Vicuna-13B2024.02 | 92 | 88.17 | |
| ParaphraseTarget Model=Vicuna-13B2024.02 | 84 | 31.69 | |
| Response CheckTarget Model=Vicuna-13B2024.02 | 30 | 82.37 | |
| No defenseTarget Model=Vicuna-13B2024.02 | 8 | 53.84 |