Evaluating robustness of neural networks

Abstract

The robustness of neural networks to adversarial examples has received great attention due to security implications. Despite various attack approaches to crafting visually imperceptible adversarial examples, little has been developed towards a comprehensive metric of robustness. This thesis is dedicated to developing several robustness quantification frameworks for deep neural networks against both adversarial and non-adversarial input perturbations, including the first robustness score CLEVER, efficient certification algorithms Fast-Lin, CROWN, CNN-Cert, and probabilistic robustness verification algorithm PROVEN. Our proposed approaches are computationally efficient and provide good quality of robustness estimates and certificates as demonstrated by extensive experiments on MNIST, CIFAR and ImageNet.