Certified Accuracy and Robustness: How different architectures stand up to adversarial attacks

Azryl Elmy Sarih, Nagender Aneja, Ong Wee Hong

Abstract: Highlights•Evaluated certified accuracy of ResNet, EfficientNet, ViT, and CaiT models.•Transformers CaiT and ViT-B/16 outperform CNNs in adversarial robustness.•Certified accuracy is boosted with adversarial training.•CaiT achieves strong robustness with low abstention, even without adversarial training.•Architectural design matters more than depth or width.