This document discusses adversarial attacks and defenses in deep learning models from an interpretation perspective. It categorizes interpretation strategies into feature-level interpretation and model-level interpretation. Feature-level interpretation techniques like gradient-based methods and influence functions can help understand adversarial attacks. Model-level interpretation of components and representations can also aid in attacking models. Additionally, feature and model-level interpretation can assist in developing defenses through techniques like model robustification, adversarial detection, and representation interpretation. The document outlines algorithms and methodologies for interpreting adversarial machine learning and considers challenges in interpreting adversarial examples.