Targeted Refusal on Targeted Refusal LLaMA2-7B-Chat (test)
100BadNetsFT
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FTThreat Model=Full Params2026.03 | 100 | 76.28 | 68.46 | 94.42 | 74.15 | 82.66 | |
| No DefenseThreat Model=N/A2026.03 | 98.94 | 73.99 | 63.31 | 95.83 | 77.98 | 82.01 | |
| PruningThreat Model=Lora Adapter2026.03 | 94.5 | 74.78 | 62.45 | 93.33 | 73.86 | 79.78 | |
| Fine-PruningThreat Model=Lora Adapter2026.03 | 90.67 | 52.66 | 48.34 | 82.31 | 53.04 | 65.36 | |
| PruningThreat Model=Full Params2026.03 | 84.68 | 39.52 | 55.58 | 86.88 | 62.37 | 65.81 | |
| QuantizationThreat Model=Full Params2026.03 | 68.32 | 32.84 | 18.29 | 64.02 | 34.33 | 43.56 | |
| Fine-PruningThreat Model=Full Params2026.03 | 59.09 | 27.62 | 36.84 | 56.33 | 48.32 | 45.64 | |
| FTThreat Model=Lora Adapter2026.03 | 25.18 | 44.56 | 42.38 | 84.37 | 65.23 | 52.34 | |
| CROWThreat Model=Full Params2026.03 | 21.93 | 43.33 | 40.53 | 88.66 | 62.57 | 51.4 | |
| BD-VAXThreat Model=Lora Adapter2026.03 | 10.66 | 8.24 | 12.32 | 9.37 | 13.22 | 10.76 | |
| BD-VAXThreat Model=Full Params2026.03 | 7.54 | 5.56 | 8.43 | 5.32 | 6.5 | 6.67 |