Skip to main content

TableĀ 4 Test accuracies showing the impact of adversarial training, clean refers to the original testing data, FGSM refers to adversary examples derived from Fast Gradient Sign Method and PGD refers to adversarial examples derived from Projected Gradient Descent [83]

From: A survey on Image Data Augmentation for Deep Learning

Models MNIST CIFAR-10
Clean FGSM PGD Clean FGSM PGD
Standard 0.9939 0.0922 0 0.9306 0.5524 0.0256
Adversarially trained 0.9932 0.9492 0.0612 0.8755 0.8526 0.1043
Our method 0.9903 0.9713 0.9171 0.8714 0.6514 0.3440