Robust deep learning

Abstract

Machine Learning models are deployed in various tasks including image classification, malware detection, network intrusion detection, etc. But recent work has demonstrated that even state-of-the-art deep neural networks, which excel at such tasks, are vulnerable to a class of malicious inputs known as Adversarial Examples. These examples are non-random inputs that are almost indistinguishable from natural data and yet are classified incorrectly. In this report, I try to explain the reason for existence of adversarial examples, discuss some of the various attacks developed to exploit the weaknesses of deep neural networks over the years and provide an analysis of such attacks over a subset of visually distinct classes of ImageNet. We then move onto the layer-wise analysis of the network developed by us, and also discuss similar works done by people in the context of adversarial attacks.