Beyond Adversarial Training: Min-Max Optimization in Adversarial Attack and Defense