Question Answering on SQuAD v2 (Utility/ASR)
1ASR ScoreNo Defense
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| No DefenseAttack=Strategy2026.04 | 1 | 0.32 | |
| PromptArmorAttack=Strategy2026.04 | 1 | 0.36 | |
| PromptGuardAttack=Strategy2026.04 | 1 | — | |
| No DefenseAttack=Combined2026.04 | 0.97 | 0.52 | |
| DataFilterAttack=Strategy2026.04 | 0.93 | 0.38 | |
| No DefenseAttack=Direct2026.04 | 0.86 | 0.56 | |
| PISanitizerAttack=Strategy2026.04 | 0.85 | 0.48 | |
| DataSentinelAttack=Strategy2026.04 | 0.78 | — | |
| PromptArmorAttack=Direct2026.04 | 0.77 | 0.66 | |
| DataFilterAttack=Direct2026.04 | 0.74 | 0.65 | |
| PIGuardAttack=Strategy2026.04 | 0.71 | — | |
| PromptArmorAttack=Combined2026.04 | 0.6 | 0.74 | |
| DataSentinelAttack=Direct2026.04 | 0.47 | — | |
| DataFilterAttack=Combined2026.04 | 0.24 | 0.69 | |
| PromptGuardAttack=Direct2026.04 | 0.24 | — | |
| DataSentinelAttack=Combined2026.04 | 0.15 | — | |
| PIGuardAttack=Direct2026.04 | 0.15 | — | |
| SecAlign++Attack=Strategy2026.04 | 0.09 | 0.91 | |
| PISanitizerAttack=Direct2026.04 | 0.04 | 0.95 | |
| PISanitizerAttack=Combined2026.04 | 0.01 | 0.95 | |
| SecAlign++Attack=No Attack2026.04 | 0.01 | 0.84 | |
| SecAlign++Attack=Combined2026.04 | 0.01 | 0.78 | |
| SecAlign++Attack=Direct2026.04 | 0.01 | 0.82 | |
| DataFilterAttack=No Attack2026.04 | 0.01 | 0.99 | |
| No DefenseAttack=No Attack2026.04 | 0 | 1 | |
| PISanitizerAttack=No Attack2026.04 | 0 | 0.99 | |
| PromptArmorAttack=No Attack2026.04 | 0 | 1 | |
| DataSentinelAttack=No Attack2026.04 | 0 | 0.99 | |
| PromptGuardAttack=No Attack2026.04 | 0 | 0.96 | |
| PromptGuardAttack=Combined2026.04 | 0 | — | |
| Attn.TrackerAttack=No Attack2026.04 | 0 | 0.61 | |
| Attn.TrackerAttack=Combined2026.04 | 0 | — | |
| Attn.TrackerAttack=Direct2026.04 | 0 | — | |
| Attn.TrackerAttack=Strategy2026.04 | 0 | — | |
| PIGuardAttack=No Attack2026.04 | 0 | 1 | |
| PIGuardAttack=Combined2026.04 | 0 | — |