Uniqueness of the Adversarial Bayes Classifier and its Implications for Consistent Adversarial Training

The adversarial Bayes classifier is unique up to degeneracy if and only if convex surrogate losses are adversarially consistent for the given data distribution.