Backdoor Detection on Simple IHU Llama 8B
0.992AUROCFunctional Attribution
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Functional AttributionNum. samples=40962026.04 | 0.992 | 100 | 99 | |
| Functional AttributionNum. samples=20482026.04 | 0.988 | 100 | 88 | |
| Functional AttributionNum. samples=10242026.04 | 0.987 | 98 | 48 | |
| VAENum. samples=40962026.04 | 0.976 | 100 | 99 | |
| Maha++Num. samples=40962026.04 | 0.974 | 100 | 99 | |
| MahaNum. samples=40962026.04 | 0.973 | 100 | 99 | |
| Maha++Num. samples=20482026.04 | 0.948 | 100 | 88 | |
| MahaNum. samples=20482026.04 | 0.943 | 100 | 88 | |
| VAENum. samples=20482026.04 | 0.832 | 100 | 88 | |
| TEDNum. samples=40962026.04 | 0.83 | 100 | 99 | |
| Maha++Num. samples=10242026.04 | 0.825 | 98 | 48 | |
| MahaNum. samples=10242026.04 | 0.79 | 98 | 48 | |
| TEDNum. samples=20482026.04 | 0.731 | 100 | 88 | |
| TEDNum. samples=10242026.04 | 0.682 | 98 | 48 | |
| VAENum. samples=10242026.04 | 0.455 | 98 | 48 |