Prompt Injection Defense on Jailbreak MI-FGSM
4Attack Success RateGUARD
Evaluation Results
| Method | Links | |
|---|---|---|
| GUARDPerturbation limit (ε)=162026.05 | 4 | |
| Gradient-guided Token SuppressionPerturbation limit (ε)=162026.05 | 4 | |
| Gradient-guided Token SuppressionPerturbation limit (ε)=322026.05 | 5 | |
| CIDERPerturbation limit (ε)=162026.05 | 6 | |
| GUARDPerturbation limit (ε)=322026.05 | 7 | |
| CIDERPerturbation limit (ε)=322026.05 | 16 | |
| ECSOPerturbation limit (ε)=162026.05 | 17 | |
| ECSOPerturbation limit (ε)=322026.05 | 18 | |
| ASTRAPerturbation limit (ε)=322026.05 | 33 | |
| ASTRAPerturbation limit (ε)=162026.05 | 42 | |
| W/O DEFENSEPerturbation limit (ε)=322026.05 | 95 | |
| W/O DEFENSEPerturbation limit (ε)=162026.05 | 96 |