Robust Image Classification on CIFAR-10 (test) (PGD Attack Metrics)
94.8Accuracy (Clean)Normal Training
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Normal TrainingBackbone=Wide Residual Network 40-2, Perturbation budget (epsilon)=8.0/255, Batch size=128, Initial learning rate=0.1, Weight decay=5e-4, Data augmentation=random cropping and mirroring2019.06 | 94.8 | 0 | 0 | |
| Adversarial TrainingBackbone=Wide Residual Network 40-2, Perturbation budget (epsilon)=8.0/255, Batch size=128, Initial learning rate=0.1, Weight decay=5e-4, Data augmentation=random cropping and mirroring2019.06 | 84.2 | 44.8 | 44.8 | |
| Adversarial Training + Auxiliary RotationsBackbone=Wide Residual Network 40-2, Perturbation budget (epsilon)=8.0/255, Batch size=128, Initial learning rate=0.1, Weight decay=5e-4, Data augmentation=random cropping and mirroring2019.06 | 83.5 | 50.4 | 50.4 |