Backdoor Rectification on ImageNet
69.05Overall AccuracyTrojaned model
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Trojaned model2026.03 | 69.05 | 87.24 | |
| Dyn. rectifyingnumber of samples (n)=202026.03 | 68.84 | 0.12 | |
| Fine-tuned modelnumber of samples (n)=102026.03 | 68.66 | 33.73 | |
| Fine-tuned modelnumber of samples (n)=202026.03 | 68.42 | 21.86 | |
| Stat. rectifyingnumber of samples (n)=12026.03 | 67.87 | 3.01 | |
| A-ClArCnumber of samples (n)=202026.03 | 67.17 | 8.73 | |
| Dyn. rectifyingnumber of samples (n)=12026.03 | 66.77 | 1.61 | |
| Fine-tuned modelnumber of samples (n)=12026.03 | 65.95 | 79.91 | |
| Patched modelnumber of samples (n)=202026.03 | 65.59 | 13.81 | |
| P-ClArCnumber of samples (n)=202026.03 | 65.42 | 8.09 |