Statistically Robust Neural Network Classification

Benjie Wang, Stefan Webb, Tom Rainforth

Despite their numerous successes, there are many scenarios where adversarial risk metrics do not provide an appropriate measure of robustness. For example, test-time perturbations…