Enhancing Robust Representation in Adversarial Training: Alignment and Exclusion Criteria