Provable Robustness of Adversarial Training
                for Learning Halfspaces with Noise

                   Difan Zou * 1 Spencer Frei * 2 Quanquan Gu 1

             Abstract                To formalize the above comment, let us define the ro-
   We analyze the properties of adversarial train-       bust error of a classifier. Let D be a distribution over
   ing for learning adversarially robust halfspaces      (x, y) ∈ Rd × {±1}, and let f : Rd → {±1} be a hy-
   in the presence of agno ...