Negative Sentiment Backdoor Detection on Gemma 2 9B
0Attack Success Rate (ASR)LCF
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LCFAttack=BadNet2026.04 | 0 | — | 15 | |
| LCFAttack=Sleeper2026.04 | 0 | — | 10.5 | |
| LCFAttack=VPI2026.04 | 0 | — | 13.5 | |
| LCFAttack=CTBA2026.04 | 0 | — | 13 | |
| LCFAttack=MTBA2026.04 | 0 | — | 19 | |
| LCFAttack=StyleBkd2026.04 | 0 | — | 14 | |
| LCFAttack=SynBkd2026.04 | 0 | — | 20.5 | |
| LCFAttack=Mean (all 14)2026.04 | 0.7 | — | 14.8 | |
| CleanGenAttack=SynBkd2026.04 | 6.5 | 2.5 | — | |
| CleanGenAttack=MTBA2026.04 | 12 | 1.5 | — | |
| CleanGenAttack=Mean (all 14)2026.04 | 12.3 | 1.6 | — | |
| CleanGenAttack=Sleeper2026.04 | 14.5 | 1.5 | — | |
| CleanGenAttack=VPI2026.04 | 16.5 | 2 | — | |
| CleanGenAttack=BadNet2026.04 | 26 | 2 | — | |
| CleanGenAttack=StyleBkd2026.04 | 26.7 | 1 | — | |
| CleanGenAttack=CTBA2026.04 | 32 | 1.5 | — | |
| DecodingAttack=MTBA2026.04 | 82 | 2.5 | — | |
| DecodingAttack=SynBkd2026.04 | 90 | 1.5 | — | |
| DecodingAttack=CTBA2026.04 | 91.5 | 1 | — | |
| DecodingAttack=BadNet2026.04 | 92.5 | 1.5 | — | |
| DecodingAttack=Mean (all 14)2026.04 | 92.7 | 1.9 | — | |
| DecodingAttack=VPI2026.04 | 93.5 | 2 | — | |
| DecodingAttack=StyleBkd2026.04 | 93.6 | 10 | — | |
| DecodingAttack=Sleeper2026.04 | 97.5 | 1 | — | |
| PruningAttack=Sleeper2026.04 | 98.5 | 1.5 | — | |
| PruningAttack=SynBkd2026.04 | 99 | 2 | — | |
| PruningAttack=VPI2026.04 | 99.5 | 1.5 | — | |
| PruningAttack=Mean (all 14)2026.04 | 99.6 | 3.8 | — | |
| Fine-tuneAttack=Mean (all 14)2026.04 | 99.8 | 3 | — | |
| CROWAttack=Mean (all 14)2026.04 | 99.9 | 3.4 | — | |
| Fine-tuneAttack=BadNet2026.04 | 100 | 1.5 | — | |
| CROWAttack=BadNet2026.04 | 100 | 1.5 | — | |
| PruningAttack=BadNet2026.04 | 100 | 2 | — | |
| Fine-tuneAttack=Sleeper2026.04 | 100 | 1.5 | — | |
| CROWAttack=Sleeper2026.04 | 100 | 1.5 | — | |
| Fine-tuneAttack=VPI2026.04 | 100 | 1 | — | |
| CROWAttack=VPI2026.04 | 100 | 1.5 | — | |
| Fine-tuneAttack=CTBA2026.04 | 100 | 2.5 | — | |
| CROWAttack=CTBA2026.04 | 100 | 1 | — | |
| PruningAttack=CTBA2026.04 | 100 | 2.5 | — | |
| Fine-tuneAttack=MTBA2026.04 | 100 | 3.5 | — | |
| CROWAttack=MTBA2026.04 | 100 | 3.5 | — | |
| PruningAttack=MTBA2026.04 | 100 | 4 | — | |
| Fine-tuneAttack=StyleBkd2026.04 | 100 | 17 | — | |
| CROWAttack=StyleBkd2026.04 | 100 | 24.5 | — | |
| PruningAttack=StyleBkd2026.04 | 100 | 20 | — | |
| Fine-tuneAttack=SynBkd2026.04 | 100 | 1.5 | — | |
| CROWAttack=SynBkd2026.04 | 100 | 2 | — |