RobustPre-Trainingby AdversarialContrastiveLearning

The labeling scarcity is even amplified when we come to adversarially robust deep learning [9], i.e., to training deep models that are not fooled by maliciously crafted, although imperceivable perturbations.