Robustness to Adversarial Examples on SVHN (test)
72.8Accuracy (FGSM)PGD training (7-steps)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| PGD training (7-steps)Backbone=WideResNet20-10, Training protocol=adversarial training2018.06 | 72.8 | — | — | — | — | — | |
| None (clean)Defense=PGD-AT, Defense budget=8/255, Backbone=ResNet-182022.01 | 71.83 | 93.95 | 57.15 | 56.02 | 54.93 | 50.5 | |
| Manifold MixupBackbone=WideResNet20-10, Alpha (α)=22018.06 | 65.91 | — | — | — | — | — | |
| Hypocritical PerturbationDefense=PGD-AT, Defense budget=8/255, Backbone=ResNet-182022.01 | 59.41 | 96.06 | 38.17 | 37.29 | 40.54 | 35.43 | |
| Input MixupBackbone=WideResNet20-10, Alpha (α)=12018.06 | 56.98 | — | — | — | — | — | |
| Adversarial PoisoningDefense=PGD-AT, Defense budget=8/255, Backbone=ResNet-182022.01 | 56.12 | 87.5 | 46.71 | 46.32 | 45.7 | 42.48 | |
| No MixupBackbone=WideResNet20-102018.06 | 21.49 | — | — | — | — | — |