Local approximations of deep learning models for black-box adversarial attacks

Abstract

We study the problem of generating adversarial examples for image classifiers in the black-box setting (when the model is available only as an oracle). We unify two seemingly orthogonal and concurrent lines of work in black-box adversarial generation: query-based attacks and substitute models. In particular, we reinterpret adversarial transferability as a strong gradient prior. Based on this unification, we develop a method for integrating model-based priors into the generation of black-box attacks. The resulting algorithms significantly improve upon the current state-of-the-art in black-box adversarial attacks across a wide range of threat models.