Robust Learnability of Real-Valued Functions under Adversarial Perturbations

Real-valued function classes with finite fat-shattering dimension are learnable in both realizable and agnostic settings under adversarial perturbations, with proper learning algorithms for convex function classes.