Adversarial Image Classification on CIFAR-10 (test) (AutoAttack Metrics)
88.9Clean AccuracyDAJAT (Base, 2*AA)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DAJAT (Base, 2*AA)Steps=2+4, Backbone=WideResNet-342022.10 | 88.9 | 57.22 | 56.96 | |
| DAJAT (Base, 3*AA)Steps=2+6, Backbone=WideResNet-342022.10 | 88.64 | 57.34 | 57.05 | |
| DAJAT (Base, AA)Steps=2+2, Backbone=WideResNet-342022.10 | 87.87 | 56.97 | 56.68 | |
| ACAT (Base, 2step)Steps=2, Backbone=WideResNet-342022.10 | 86.71 | 55.58 | 55.36 | |
| DAJAT (Base, 3*AA)Steps=2+6, Backbone=ResNet-18, Time/epoch (sec)=2802022.10 | 86.67 | 51.81 | 51.56 | |
| NuAT2-WASteps=2, Backbone=WideResNet-342022.10 | 86.32 | 55.08 | 54.76 | |
| PGD-ATSteps=10, Backbone=WideResNet-342022.10 | 86.07 | 52.7 | 52.19 | |
| DAJAT (Base, 2*AA)Steps=2+4, Backbone=ResNet-18, Time/epoch (sec)=2192022.10 | 85.99 | 51.71 | 51.48 | |
| DAJAT (Base, AA)Steps=2+2, Backbone=ResNet-18, Time/epoch (sec)=1602022.10 | 85.6 | 51.27 | 51.06 | |
| TRADES-AWP-WA (200 epochs)Steps=10, Backbone=WideResNet-34, Epochs=2002022.10 | 85.36 | 56.35 | 56.17 | |
| TRADES-AWPSteps=10, Backbone=WideResNet-342022.10 | 85.19 | 55.87 | 55.69 | |
| TRADES-AWP-WASteps=10, Backbone=WideResNet-342022.10 | 85.1 | 56.07 | 55.87 | |
| ACAT (Base, 2step)Steps=2, Backbone=ResNet-18, Time/epoch (sec)=952022.10 | 82.41 | 50 | 49.8 | |
| NuAT2-WASteps=2, Backbone=ResNet-18, Time/epoch (sec)=1092022.10 | 82.21 | 50.97 | 50.75 | |
| TRADES-AWP-WA (200 epochs)Steps=10, Backbone=ResNet-18, Time/epoch (sec)=228, Epochs=2002022.10 | 81.99 | 51.65 | 51.45 | |
| PGD-ATSteps=10, Backbone=ResNet-18, Time/epoch (sec)=1822022.10 | 81.12 | 49.08 | 48.75 | |
| TRADES-AWPSteps=10, Backbone=ResNet-18, Time/epoch (sec)=2282022.10 | 80.47 | 50.06 | 49.87 | |
| TRADES-AWP-WASteps=10, Backbone=ResNet-18, Time/epoch (sec)=2282022.10 | 80.41 | 49.89 | 49.67 |