Backdoor Detection on Simple IHU Gemma 2B
1AUROCFunctional Attribution
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Functional AttributionNum. samples=40962026.04 | 1 | 100 | 98 | |
| Functional AttributionNum. samples=10242026.04 | 0.999 | 90 | 74 | |
| Functional AttributionNum. samples=20482026.04 | 0.999 | 94 | 98 | |
| TEDNum. samples=40962026.04 | 0.972 | 100 | 98 | |
| VAENum. samples=40962026.04 | 0.956 | 100 | 98 | |
| Maha++Num. samples=40962026.04 | 0.953 | 100 | 98 | |
| MahaNum. samples=40962026.04 | 0.943 | 100 | 98 | |
| TEDNum. samples=20482026.04 | 0.932 | 94 | 98 | |
| Maha++Num. samples=20482026.04 | 0.928 | 94 | 98 | |
| MahaNum. samples=20482026.04 | 0.903 | 94 | 98 | |
| Maha++Num. samples=10242026.04 | 0.834 | 90 | 74 | |
| VAENum. samples=20482026.04 | 0.801 | 94 | 98 | |
| MahaNum. samples=10242026.04 | 0.779 | 90 | 74 | |
| TEDNum. samples=10242026.04 | 0.77 | 90 | 74 | |
| VAENum. samples=10242026.04 | 0.544 | 90 | 74 |