List your company

Negative Sentiment Backdoor Detection on Gemma 2 9B

0Attack Success Rate (ASR)LCF
-4235077Apr 27, 2026

Evaluation results

MethodLinks
LCFAttack=BadNet2026.040—15
LCFAttack=Sleeper2026.040—10.5
LCFAttack=VPI2026.040—13.5
LCFAttack=CTBA2026.040—13
LCFAttack=MTBA2026.040—19
LCFAttack=StyleBkd2026.040—14
LCFAttack=SynBkd2026.040—20.5
LCFAttack=Mean (all 14)2026.040.7—14.8
CleanGenAttack=SynBkd2026.046.52.5—
CleanGenAttack=MTBA2026.04121.5—
CleanGenAttack=Mean (all 14)2026.0412.31.6—
CleanGenAttack=Sleeper2026.0414.51.5—
CleanGenAttack=VPI2026.0416.52—
CleanGenAttack=BadNet2026.04262—
CleanGenAttack=StyleBkd2026.0426.71—
CleanGenAttack=CTBA2026.04321.5—
DecodingAttack=MTBA2026.04822.5—
DecodingAttack=SynBkd2026.04901.5—
DecodingAttack=CTBA2026.0491.51—
DecodingAttack=BadNet2026.0492.51.5—
DecodingAttack=Mean (all 14)2026.0492.71.9—
DecodingAttack=VPI2026.0493.52—
DecodingAttack=StyleBkd2026.0493.610—
DecodingAttack=Sleeper2026.0497.51—
PruningAttack=Sleeper2026.0498.51.5—
PruningAttack=SynBkd2026.04992—
PruningAttack=VPI2026.0499.51.5—
PruningAttack=Mean (all 14)2026.0499.63.8—
Fine-tuneAttack=Mean (all 14)2026.0499.83—
CROWAttack=Mean (all 14)2026.0499.93.4—
Fine-tuneAttack=BadNet2026.041001.5—
CROWAttack=BadNet2026.041001.5—
PruningAttack=BadNet2026.041002—
Fine-tuneAttack=Sleeper2026.041001.5—
CROWAttack=Sleeper2026.041001.5—
Fine-tuneAttack=VPI2026.041001—
CROWAttack=VPI2026.041001.5—
Fine-tuneAttack=CTBA2026.041002.5—
CROWAttack=CTBA2026.041001—
PruningAttack=CTBA2026.041002.5—
Fine-tuneAttack=MTBA2026.041003.5—
CROWAttack=MTBA2026.041003.5—
PruningAttack=MTBA2026.041004—
Fine-tuneAttack=StyleBkd2026.0410017—
CROWAttack=StyleBkd2026.0410024.5—
PruningAttack=StyleBkd2026.0410020—
Fine-tuneAttack=SynBkd2026.041001.5—
CROWAttack=SynBkd2026.041002—