Attribution Adversarial Sensitivity on Imagenette 224x224 (test)
0.61VGradNo Regularization (Base)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| No Regularization (Base)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0.61 | — | 0.1 | 0.22 | 0.1 | 0.03 | |
| Sharp. Aware Min. (SAM)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0.3 | — | 0.16 | 0.12 | 0.1 | 0.04 | |
| Adv. Trainig Reg. (ATR)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0.21 | — | 0.04 | 0.15 | 0.05 | 0.03 | |
| Explicit Curv. Reg. (ECR)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0.11 | — | 0.08 | 0.09 | 0.04 | 0.01 | |
| No Regularization (Base)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.11 | — | 0.04 | 0.06 | 0.01 | 0.09 | |
| Ante-hoc Act. Reg. (AAR)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0.1 | — | 0.04 | 0.05 | 0.04 | 0.03 | |
| Implicit Curv. Reg. (ICR)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0.1 | — | 0.05 | 0.05 | 0.04 | 0.01 | |
| Sharp. Aware Min. (SAM)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.1 | — | 0.04 | 0.05 | 0.01 | 0.1 | |
| Explicit Curv. Reg. (ECR)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.08 | — | 0.05 | 0.09 | 0.03 | 0.06 | |
| Adv. Trainig Reg. (ATR)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.08 | — | 0.02 | 0.05 | 0 | 0.06 | |
| Ante-hoc Act. Reg. (AAR)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.07 | — | 0.03 | 0.04 | 0.01 | 0.07 | |
| Implicit Curv. Reg. (ICR†)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.07 | — | 0.04 | 0.03 | 0.02 | 0.07 | |
| Post-hoc Act. Reg. (PAR)Architecture=ViT-B/16, Input Resolution=224x2242026.05 | 0.06 | — | 0.02 | 0.05 | 0 | 0.06 | |
| Implicit Curv. Reg. (ICR†)Architecture=ViT-B/16 (Kernelized), Input Resolution=224x2242026.05 | 0.05 | — | 0.02 | 0.05 | 0.02 | 0.05 | |
| Post-hoc Act. Reg. (PAR)Architecture=ResNet50, Input Resolution=224x2242026.05 | 0 | — | 0 | 0 | 0 | 0 |