DLR: Adversarial examples detection and label recovery for deep neural networks

Keji Han, Yao Ge, Ruchuan Wang, Yun Li

Abstract: Highlights•We propose DLR to recover ground-truth labels of detected adversarial examples.•We analyze the potential of separate modules for legitimate and adversarial examples.•We analyze adversarial robustness of a multi-class autoencoder-based classifier.