Sentiment Steering on LLaMA2-7B-Chat (test)
2.01BadNetsBD-VAX
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BD-VAXThreat Model=Lora Adapter2026.03 | 2.01 | 0.51 | 0 | 0 | 2 | 0.91 | |
| BD-VAXThreat Model=Full Params2026.03 | 2.51 | 1.01 | 0 | 0.5 | 6.5 | 2.1 | |
| Fine-PruningThreat Model=Lora Adapter2026.03 | 13.57 | 3.53 | 0 | 0 | 13.5 | 6.12 | |
| Fine-PruningThreat Model=Full Params2026.03 | 18.59 | 1.01 | 0.51 | 1.01 | 29.5 | 10.94 | |
| CROWThreat Model=Full Params2026.03 | 21.11 | 3.08 | 0.5 | 0.5 | 18.09 | 8.66 | |
| FTThreat Model=Lora Adapter2026.03 | 23.59 | 0 | 0 | 3.01 | 24.5 | 10.62 | |
| QuantizationThreat Model=Full Params2026.03 | 31.5 | 5 | 2 | 4 | 39.29 | 16.36 | |
| PruningThreat Model=Full Params2026.03 | 36.3 | 4 | 1.51 | 4.5 | 20.6 | 13.78 | |
| PruningThreat Model=Lora Adapter2026.03 | 47.47 | 9.02 | 2.53 | 2.08 | 50.48 | 22.32 | |
| No DefenseThreat Model=N/A2026.03 | 59.3 | 13.68 | 4.3 | 3.52 | 60 | 28.16 | |
| FTThreat Model=Full Params2026.03 | 60 | 13.75 | 5.08 | 3.52 | 63.47 | 29.96 |