On the Coarse Robustness of Classifiers
Ismail Alkhouri, Stanley Bak, Alvaro Velasquez, George Atia · 2022
Standard measures of robustness, derived from the least amount of adversarial perturbation, often fail to gauge the ability of a classifier to recognize the coarse genres. It is desirable to have a classifier with high coarse robustness with respect to a grouping that is consistent with the class semantics, so that semantically-plausible coarse categories remain invariant to imperceptible perturbations. In this work, we formalize a new notion of coarse robustness that is defined with respect to a specified grouping of the class labels. We formulate an optimization problem to obtain the optimal grouping, and develop an algorithm that is shown to perform on par with brute force search. Moreover, we propose a training mechanism that incorporates the coarse label information in addition to the finer ones. We empirically and theoretically show that this mechanism improves the proposed coarse notion of robustness while only requiring a relatively small additional parameters and training time.