End-to-End Defense in RAG on SciFact
0ASRTrustRAG
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TrustRAGAttack Scenario=GL2026.05 | 0 | 0 | |
| PromptGuardAttack Scenario=GL2026.05 | 0 | 1 | |
| CleanBaseAttack Scenario=GL2026.05 | 0 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=GL2026.05 | 0 | 0 | |
| Cordon-MASSeed=422026.05 | 2 | — | |
| CleanBaseAttack Scenario=PT2026.05 | 3 | 0.4 | |
| CleanBase + MetaSecAlignAttack Scenario=PT2026.05 | 3 | 0.4 | |
| PromptGuardAttack Scenario=FC2026.05 | 4 | 0 | |
| CleanBaseAttack Scenario=MA2026.05 | 4 | 0 | |
| PromptGuardAttack Scenario=NA2026.05 | 5 | 0 | |
| CleanBaseAttack Scenario=CI2026.05 | 5 | 0 | |
| CleanBaseAttack Scenario=FC2026.05 | 6 | 0 | |
| PromptGuardAttack Scenario=MA2026.05 | 7 | 0 | |
| CleanBaseAttack Scenario=PB2026.05 | 7 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=NA2026.05 | 7 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=FC2026.05 | 7 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=MA2026.05 | 7 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=PB2026.05 | 7 | 0 | |
| CleanBaseAttack Scenario=NA2026.05 | 8 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=CI2026.05 | 8 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=CA2026.05 | 8 | 0 | |
| PromptGuardAttack Scenario=CI2026.05 | 9 | 0 | |
| CleanBaseAttack Scenario=CA2026.05 | 9 | 0 | |
| PromptGuardAttack Scenario=CA2026.05 | 10 | 0 | |
| TrustRAGAttack Scenario=MA2026.05 | 12 | 0 | |
| CleanBase + MetaSecAlignAttack Scenario=PW2026.05 | 12 | 5.4 | |
| TrustRAGAttack Scenario=NA2026.05 | 14 | 100 | |
| CleanBaseAttack Scenario=PW2026.05 | 15 | 5.4 | |
| TrustRAGAttack Scenario=CI2026.05 | 16 | 0 | |
| TrustRAGAttack Scenario=PB2026.05 | 16 | 0 | |
| TrustRAGAttack Scenario=FC2026.05 | 17 | 0 | |
| TrustRAGAttack Scenario=PT2026.05 | 17 | 0 | |
| TrustRAGAttack Scenario=PW2026.05 | 18 | 0.4 | |
| TrustRAGAttack Scenario=CA2026.05 | 20 | 0 | |
| MetaSecAlignAttack Scenario=FC2026.05 | 21 | 100 | |
| DataSentinelAttack Scenario=CA2026.05 | 21 | 9.6 | |
| PromptGuardAttack Scenario=PT2026.05 | 22 | 33.2 | |
| MetaSecAlignAttack Scenario=NA2026.05 | 23 | 100 | |
| MetaSecAlignAttack Scenario=PB2026.05 | 23 | 100 | |
| MetaSecAlignAttack Scenario=MA2026.05 | 26 | 100 | |
| MetaSecAlignAttack Scenario=CI2026.05 | 29 | 100 | |
| MetaSecAlignAttack Scenario=PW2026.05 | 29 | 100 | |
| MetaSecAlignAttack Scenario=CA2026.05 | 30 | 100 | |
| MetaSecAlignAttack Scenario=PT2026.05 | 30 | 98.8 | |
| DataSentinelAttack Scenario=PT2026.05 | 33 | 98.8 | |
| No DefenseAttack Scenario=PT2026.05 | 34 | 98.8 | |
| DataSentinelAttack Scenario=CI2026.05 | 38 | 20.4 | |
| DataSentinelAttack Scenario=FC2026.05 | 38 | 18.8 | |
| DebateSeed=422026.05 | 38 | — | |
| DataSentinelAttack Scenario=MA2026.05 | 40 | 18.6 | |
| RobustRAGSeed=422026.05 | 44 | — | |
| PromptGuardAttack Scenario=PB2026.05 | 47 | 62.4 | |
| No DefenseAttack Scenario=PB2026.05 | 48 | 100 | |
| PromptGuardAttack Scenario=PW2026.05 | 54 | 70.2 | |
| MetaSecAlignAttack Scenario=GL2026.05 | 55 | 100 | |
| DataSentinelAttack Scenario=PB2026.05 | 56 | 100 | |
| ParaphraseSeed=422026.05 | 58 | — | |
| TrustRAGSeed=422026.05 | 60 | — | |
| No DefenseAttack Scenario=PW2026.05 | 62 | 100 | |
| Vanilla RAGSeed=422026.05 | 62 | — | |
| DataSentinelAttack Scenario=NA2026.05 | 65 | 66 | |
| DataSentinelAttack Scenario=PW2026.05 | 67 | 99.2 | |
| DataSentinelAttack Scenario=GL2026.05 | 80 | 51 | |
| No DefenseAttack Scenario=FC2026.05 | 89 | 100 | |
| No DefenseAttack Scenario=GL2026.05 | 90 | 100 | |
| No DefenseAttack Scenario=CA2026.05 | 97 | 100 | |
| No DefenseAttack Scenario=NA2026.05 | 99 | 100 | |
| No DefenseAttack Scenario=CI2026.05 | 99 | 100 | |
| No DefenseAttack Scenario=MA2026.05 | 99 | 100 |