关键词: 机器学习安全, 深度学习安全, 攻防竞赛, 对抗攻击, 成员推理攻击, 隐私保护

Abstract: Artificial intelligence has penetrated into every corners of our life and brought humans great convenience. Especially in recent years, with the vigorous development of the deep learning branch in machine learning, there are more and more related applications in our life. Unfortunately, machine learning systems are suffering from many security hazards. Even worse, the popularity of machine learning systems further magnifies these hazards. In order to unveil these security hazards and assist in implementing a robust machine learning system, we conduct a comprehensive investigation of the mainstream deep learning systems. In the beginning of the study, we devise an analytical model for dissecting deep learning systems, and define our survey scope. Our surveyed deep learning systems span across four fields-image classification, audio speech recognition, malware detection, and natural language processing. We distill four types of security hazards and manifest them in multiple dimensions such as complexity, attack success rate, and damage. Furthermore, we survey defensive techniques for deep learning systems as well as their characteristics. Finally, through the observation of these systems, we propose the practical proposals of constructing robust deep learning system.

Key words: machine learning security, deep learning security, attack and defense race, adversarial attack, membership inference attack, privacy-preserving