Malware poses a severe threat to cyber security. Attackers use malware to achieve their malicious purposes, such as unauthorized access, stealing confidential data, blackmailing, etc. Machine learning-based defense methods are applied to classify malware examples. However, such methods are vulnerable to adversarial attacks, where attackers aim to generate adversarial examples that can evade detection. Defenders also develop various approaches to enhance the robustness of malware classifiers against adversarial attacks. Both attackers and defenders evolve in the continuous confrontation of malware classification. In this paper, we firstly summarize a unified malware classification framework. Then, based on the framework, we systematically survey the Defense-Attack-Enhanced-Defense process and provide a comprehensive review of (i) machine learning-based malware classification, (ii) adversarial attacks on malware classifiers, and (iii) robust malware classification. Finally, we highlight the main challenges faced by both attackers and defenders and discuss some promising future work directions. 中文翻译： 恶意软件对网络安全构成严重威胁。攻击者使用恶意软件来实现其恶意目的，例如未经授权的访问、窃取机密数据、勒索等。基于机器学习的防御方法被应用于对恶意软件示例进行分类。但是，此类方法容易受到对抗性攻击，攻击者的目标是生成可以逃避检测的对抗性示例。防御者还开发了各种方法来增强恶意软件分类器抵御对抗性攻击的稳健性。攻击者和防御者都在恶意软件分类的不断交锋中进化。在本文中，我们首先总结了一个统一的恶意软件分类框架。然后，基于框架，我们系统地调查了 Defense-Attack-Enhanced-Defense 过程，并全面回顾了 (i) 基于机器学习的恶意软件分类，(ii) 对恶意软件分类器的对抗性攻击，以及 (iii) 强大的恶意软件分类。最后，我们强调了攻击者和防御者面临的主要挑战，并讨论了一些有前途的未来工作方向。