Addressing Neural Network Robustness with Mixup and Targeted Labeling Adversarial Training

Despite their performance, Artificial Neural Networks are not reliable enough\nfor most of industrial applications. They are sensitive to noises, rotations,\nblurs and adversarial examples. There is a need to build defenses that protect\nagainst a wide range of perturbations, covering the most traditional common\ncorruptions and adversarial examples. We propose a new data augmentation\nstrategy called M-TLAT and designed to address robustness in a broad sense. Our\napproach combines the Mixup augmentation and a new adversarial training\nalgorithm called Targeted Labeling Adversarial Training (TLAT). The idea of\nTLAT is to interpolate the target labels of adversarial examples with the\nground-truth labels. We show that M-TLAT can increase the robustness of image\nclassifiers towards nineteen common corruptions and five adversarial attacks,\nwithout reducing the accuracy on clean samples.\n