Boosting Adversarial Training Using Robust Selective Data Augmentation

Abstract Artificial neural networks are currently applied in a wide variety of fields, and they near to achieving performance similar humans many tasks. Nevertheless, vulnerable adversarial attacks the form small intentionally designed perturbation, which could lead misclassifications, making these models unusable, especially applications where security is critical. The best defense against attacks, so far, training (AT), improves model’s robustness by augmenting data with examples. In this work, we show that AT can be further improved employing neighborhood each example latent space make additional targeted augmentations data. More specifically, propose robust selective augmentation (RSDA) approach enhance AT. RSDA complements inspecting quality from perspective performing transformation operations on specific neighboring samples sample space. We evaluate MNIST CIFAR-10 datasets multiple attacks. Our experiments gives significantly better results than just both clean samples.