Refusal Backdoor Detection on Gemma-2-9B
0ASRLCF
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LCFAttack=BadNet2026.04 | 0 | — | 13 | |
| LCFAttack=Sleeper2026.04 | 0 | — | 15 | |
| LCFAttack=SynBkd2026.04 | 0.5 | — | 16 | |
| LCFAttack=VPI2026.04 | 1.5 | — | 12 | |
| LCFAttack=CTBA2026.04 | 1.5 | — | 16 | |
| LCFAttack=MTBA2026.04 | 2.5 | — | 12.5 | |
| CleanGenAttack=StyleBkd2026.04 | 2.7 | 2 | — | |
| LCFAttack=StyleBkd2026.04 | 3.8 | — | 17.5 | |
| CleanGenAttack=Sleeper2026.04 | 4 | 1.5 | — | |
| CleanGenAttack=SynBkd2026.04 | 4 | 2 | — | |
| CleanGenAttack=BadNet2026.04 | 4.5 | 1.5 | — | |
| CleanGenAttack=MTBA2026.04 | 4.5 | 1 | — | |
| CleanGenAttack=VPI2026.04 | 8.5 | 1.5 | — | |
| CleanGenAttack=CTBA2026.04 | 9.5 | 1.5 | — | |
| DecodingAttack=MTBA2026.04 | 89 | 1 | — | |
| DecodingAttack=StyleBkd2026.04 | 91.9 | 3.5 | — | |
| DecodingAttack=VPI2026.04 | 93.5 | 0.5 | — | |
| DecodingAttack=CTBA2026.04 | 94 | 0.5 | — | |
| DecodingAttack=Sleeper2026.04 | 95 | 0.5 | — | |
| DecodingAttack=SynBkd2026.04 | 95.5 | 0.5 | — | |
| Fine-tuneAttack=StyleBkd2026.04 | 97.8 | 3 | — | |
| PruningAttack=StyleBkd2026.04 | 97.8 | 5 | — | |
| DecodingAttack=BadNet2026.04 | 98 | 0 | — | |
| CROWAttack=StyleBkd2026.04 | 98.9 | 2.5 | — | |
| Fine-tuneAttack=SynBkd2026.04 | 99.5 | 2.5 | — | |
| PruningAttack=SynBkd2026.04 | 99.5 | 4.5 | — | |
| Fine-tuneAttack=BadNet2026.04 | 100 | 1.5 | — | |
| CROWAttack=BadNet2026.04 | 100 | 1 | — | |
| PruningAttack=BadNet2026.04 | 100 | 3 | — | |
| Fine-tuneAttack=Sleeper2026.04 | 100 | 1.5 | — | |
| CROWAttack=Sleeper2026.04 | 100 | 1.5 | — | |
| PruningAttack=Sleeper2026.04 | 100 | 2 | — | |
| Fine-tuneAttack=VPI2026.04 | 100 | 1 | — | |
| CROWAttack=VPI2026.04 | 100 | 1 | — | |
| PruningAttack=VPI2026.04 | 100 | 1 | — | |
| Fine-tuneAttack=CTBA2026.04 | 100 | 1.5 | — | |
| CROWAttack=CTBA2026.04 | 100 | 1.5 | — | |
| PruningAttack=CTBA2026.04 | 100 | 1.5 | — | |
| Fine-tuneAttack=MTBA2026.04 | 100 | 2 | — | |
| CROWAttack=MTBA2026.04 | 100 | 2 | — | |
| PruningAttack=MTBA2026.04 | 100 | 2 | — | |
| CROWAttack=SynBkd2026.04 | 100 | 3 | — |