Adversarial Detection on ImageNet (val)
94.88AUROC (PGD)CADet
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CADetBackbone=ResNet50, Tuned on Adv=No, Training=Contrastive2022.10 | 94.88 | 95.93 | 97.56 | 96.12 | |
| HuBackbone=ResNet50, Tuned on Adv=Yes, Training=Contrastive2022.10 | 94.8 | 95.19 | 78.18 | 89.39 | |
| FSBackbone=ResNet50, Tuned on Adv=Yes, Training=Contrastive2022.10 | 94.71 | 88 | 95.98 | 92.9 | |
| LIDBackbone=ResNet50, Tuned on Adv=Yes, Training=Supervised2022.10 | 93.13 | 91.08 | 93.55 | 92.59 | |
| MahalanobisBackbone=ResNet50, Tuned on Adv=No, Training=Supervised2022.10 | 92.71 | 91.01 | 96.92 | 93.55 | |
| LIDBackbone=ResNet50, Tuned on Adv=Yes, Training=Contrastive2022.10 | 91.94 | 89.5 | 93.56 | 91.67 | |
| FSBackbone=ResNet50, Tuned on Adv=Yes, Training=Supervised2022.10 | 91.39 | 87.77 | 95.12 | 91.43 | |
| HuBackbone=ResNet50, Tuned on Adv=Yes, Training=Supervised2022.10 | 84.31 | 84.29 | 77.95 | 82.18 | |
| MahalanobisBackbone=ResNet50, Tuned on Adv=No, Training=Contrastive2022.10 | 84.14 | 83.78 | 79.9 | 82.61 | |
| Hu + self-calibrationBackbone=ResNet50, Tuned on Adv=No, Training=Contrastive2022.10 | 75.69 | 75.74 | 69.2 | 73.54 | |
| CADetBackbone=ResNet50, Tuned on Adv=No, Training=Supervised2022.10 | 75.25 | 71.02 | 83.45 | 76.57 | |
| Hu + self-calibrationBackbone=ResNet50, Tuned on Adv=No, Training=Supervised2022.10 | 66.4 | 59.58 | 71.02 | 65.67 | |
| ODINBackbone=ResNet50, Tuned on Adv=Yes, Training=Supervised2022.10 | 62.3 | 60.29 | 68.1 | 63.56 | |
| ODINBackbone=ResNet50, Tuned on Adv=Yes, Training=Contrastive2022.10 | 59.91 | 60.23 | 64.99 | 61.71 |