Image Classification on CIFAR-100 (val) (PGD Robustness)
28.4Accuracy (PGD-20)NADAR-B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| NADAR-B2021.08 | 28.4 | 62.56 | — | — | |
| RACL2021.08 | 27.8 | — | — | — | |
| NADAR-A2021.08 | 27.77 | 61.73 | — | — | |
| FreeAT-82021.08 | 25.88 | 62.13 | — | — | |
| Free m = 8replay parameter (m)=8, Training Time (minutes)=780, Backbone=Wide-Resnet 32-102019.04 | 25.88 | 62.13 | 25.58 | — | |
| Freem=82019.04 | 25.88 | 62.13 | 25.58 | 780 | |
| Free m = 10replay parameter (m)=10, Training Time (minutes)=776, Backbone=Wide-Resnet 32-102019.04 | 25.15 | 59.27 | 24.88 | — | |
| Freem=102019.04 | 25.15 | 59.27 | 24.88 | 776 | |
| Free m = 6replay parameter (m)=6, Training Time (minutes)=791, Backbone=Wide-Resnet 32-102019.04 | 23.68 | 64.87 | 23.18 | — | |
| Freem=62019.04 | 23.68 | 64.87 | 23.18 | 791 | |
| PGD-72021.08 | 23.2 | — | — | — | |
| RobNet-large2021.08 | 23.19 | — | — | — | |
| Madry et al. (7-PGD trained)Training protocol=7-PGD, Training Time (minutes)=5157, Backbone=Wide-Resnet 32-102019.04 | 22.76 | 59.87 | 22.52 | — | |
| Madry et al. (7-PGD trained)Training=7-PGD trained2019.04 | 22.76 | 59.87 | 22.52 | 5,157 | |
| Free m = 4replay parameter (m)=4, Training Time (minutes)=767, Backbone=Wide-Resnet 32-102019.04 | 20.64 | 65.28 | 20.15 | — | |
| Freem=42019.04 | 20.64 | 65.28 | 20.15 | 767 | |
| Madry et al. (2-PGD trained)Training protocol=2-PGD, Training Time (minutes)=2053, Backbone=Wide-Resnet 32-102019.04 | 17.08 | 67.94 | 16.5 | — | |
| Madry et al. (2-PGD trained)Training=2-PGD trained2019.04 | 17.08 | 67.94 | 16.5 | 2,053 | |
| Free m = 2replay parameter (m)=2, Training Time (minutes)=816, Backbone=Wide-Resnet 32-102019.04 | 15.37 | 69.2 | 14.86 | — | |
| Freem=22019.04 | 15.37 | 69.2 | 14.86 | 816 | |
| Standard2021.08 | 0 | 78.84 | — | — | |
| NaturalTraining Time (minutes)=811, Backbone=Wide-Resnet 32-102019.04 | 0 | 78.84 | 0 | — | |
| NaturalTraining=Standard Natural Training2019.04 | 0 | 78.84 | 0 | 811 |