Adversarial Image Classification on ImageNet (Linf eps=4/255, test)
62.05AccuracyMixACM + RA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| MixACM + RABackbone=ResNet-50, Train ε=02021.11 | 62.05 | 34.93 | 33.71 | 33.63 | |
| Free-ATBackbone=ResNet-50, Train ε=4/2552021.11 | 60.21 | 32.77 | 31.88 | 31.82 | |
| MixACMBackbone=ResNet-50, Train ε=02021.11 | 59.64 | 32.7 | 31.51 | 31.39 | |
| Fast-ATBackbone=ResNet-50, Train ε=4/2552021.11 | 55.45 | — | 30.28 | — |