Image Classification on CIFAR-10 (ECE and Query-Based Robustness Analysis)
2.46ECE (%)AAA-linear
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AAA-linearBackbone=Wide-ResNet-28, Perturbation Budget (l_inf)=8/255, Defense Strategy=Adversarial Attack on Attackers (linear)2022.05 | 2.46 | 94.84 | 81.36 | 80.59 | 79.41 | 76.71 | 88.86 | 83.36 | 90.62 | 85.95 | 80.86 | 78.36 | |
| NoneBackbone=Wide-ResNet-28, Perturbation Budget (l_inf)=8/255, Defense Strategy=Undefended2022.05 | 3.52 | 94.78 | 39.38 | 0.09 | 41.14 | 0.04 | 53.04 | 3.95 | 83.42 | 12.24 | 69.86 | 41.03 | |
| RNDBackbone=Wide-ResNet-28, Perturbation Budget (l_inf)=8/255, Defense Strategy=Random Noise Defense2022.05 | 6.32 | 91.05 | 60.83 | 49.15 | 61.02 | 47.82 | 76.39 | 64.34 | 86.23 | 68.19 | 70.44 | 41.65 | |
| ATBackbone=Wide-ResNet-28, Perturbation Budget (l_inf)=8/255, Defense Strategy=Adversarial Training2022.05 | 11 | 87.02 | 78.3 | 67.44 | 78.87 | 66.79 | 84.21 | 75.85 | 85.92 | 81.01 | 83.62 | 76.25 |