deep neural networks (DNNs) are sensitive to adversarial examples, resulting
in fragile and unreliable performance in the real world. Although adversarial
training (AT) is currently one of the most effective methodologies to robustify
DNNs, it is computationally very expensive (e.g., 5