On practical robustness of machine learning systems

Abstract

We consider the importance of robustness in evaluating machine learning systems, an in particular systems involving deep learning. We consider these systems' vulnerability to adversarial examples--subtle, crafted perturbations to inputs which induce large change in output. We show that these adversarial examples are not only theoretical concern, by desigining the first 3D adversarial objects, and by demonstrating that these examples can be constructed even when malicious actors have little power. We suggest a potential avenue for building robust deep learning models by leveraging generative models.