Understanding and Improving Fast Adversarial Training against $l_0$ Bounded Perturbations | Read Paper on Bytez