Are Transformers More Robust Than CNNs?

Open paper
Benchmark Model Rank Results
adversarial-robustness-on-imagenetResNet-50 (SGD, Cosine)#1Accuracy: 77.4
adversarial-robustness-on-imagenetResNet-50 (SGD, Step)#2Accuracy: 76.9
adversarial-robustness-on-imagenetDeiT-S (AdamW, Cosine)#3Accuracy: 76.8
adversarial-robustness-on-imagenetResNet-50 (AdamW, Cosine)#4Accuracy: 76.4
adversarial-robustness-on-imagenet-aDeiT-S (AdamW, Cosine)#1Accuracy: 12.2
adversarial-robustness-on-imagenet-aResNet-50 (SGD, Cosine)#2Accuracy: 3.3
adversarial-robustness-on-imagenet-aResNet-50 (SGD, Step)#3Accuracy: 3.2
adversarial-robustness-on-imagenet-aResNet-50 (AdamW, Cosine)#4Accuracy: 3.1
adversarial-robustness-on-imagenet-cDeiT-S (AdamW, Cosine)#1mean Corruption Error (mCE): 48.0
adversarial-robustness-on-imagenet-cResNet-50 (SGD, Cosine)#2mean Corruption Error (mCE): 56.9
adversarial-robustness-on-imagenet-cResNet-50 (SGD, Step)#3mean Corruption Error (mCE): 57.9
adversarial-robustness-on-imagenet-cResNet-50 (AdamW, Cosine)#4mean Corruption Error (mCE): 59.3
adversarial-robustness-on-stylized-imagenetDeiT-S (AdamW, Cosine)#1Accuracy: 13.0
adversarial-robustness-on-stylized-imagenetResNet-50 (SGD, Cosine)#2Accuracy: 8.4
adversarial-robustness-on-stylized-imagenetResNet-50 (SGD, Step)#3Accuracy: 8.3
adversarial-robustness-on-stylized-imagenetResNet-50 (AdamW, Cosine)#4Accuracy: 8.1

Rank counts only results with a code link.