Image Classification on CIFAR-10 (Corruption and L-infinity Robustness Analysis)
0Accuracy (Clean)Normal†
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Normal†Training=Standard Training2022.02 | 0 | 14.4 | 1.4 | 8 | 3.3 | 1.6 | 100 | 100 | |
| Madry†∞Training=Adversarial Training, Constraint=l_inf2022.02 | 0 | 19.3 | 2.6 | 9.3 | 11.7 | 8.4 | 38.6 | 43.4 | |
| PAT†selfTraining=Perceptual Adversarial Training, Variant=self2022.02 | 0 | 23 | 3.3 | 11.8 | 12.6 | 5.7 | 50 | 63.4 | |
| PAT†alexTraining=Perceptual Adversarial Training, Variant=alex2022.02 | 0 | 27.7 | 4.2 | 13 | 17.2 | 10 | 41.4 | 59.8 | |
| GAT-f†Training=Generative Adversarial Training, Mode=f2022.02 | 0 | 1.6 | 1.9 | 5.7 | 2.8 | 3.3 | 48.2 | 53 | |
| GAT-fs†Training=Generative Adversarial Training, Mode=fs2022.02 | 0 | 2 | 1.6 | 5.8 | 2.2 | 3.3 | 43.2 | 49 | |
| Normal∗Training=Standard Training2022.02 | 0 | 19.7 | 1.8 | 7.6 | 5.3 | 2.9 | 100 | 100 | |
| Trades∗∞Training=TRADES, Constraint=l_inf2022.02 | 0 | 23 | 2.6 | 9.5 | 18.2 | 16.9 | 34.3 | 38.2 | |
| FAT∗∞Training=Fast Adversarial Training, Constraint=l_inf2022.02 | 0 | 22.2 | 3.1 | 13.3 | 17.3 | 13.4 | 37.9 | 41.5 | |
| AWP∗∞Training=Adversarial Weight Perturbation, Constraint=l_inf2022.02 | 0 | 21.5 | 2.8 | 10.5 | 20.3 | 17.3 | 30.4 | 34.2 | |
| GAT-f∗Training=Generative Adversarial Training, Mode=f2022.02 | 0 | 1.7 | 2 | 5.6 | 2.3 | 3.1 | 47.7 | 52 | |
| GAT-fs∗Training=Generative Adversarial Training, Mode=fs2022.02 | 0 | 2.3 | 1.9 | 5.9 | 2.6 | 3.2 | 43.3 | 49.2 |