机器学习成员推理攻击研究进展与挑战

Research Progress and Challenges of Membership Inference Attacks in Machine Learning

摘要: 成员推理攻击通过对机器学习模型进行攻击可推断目标数据是否为训练数据集的成员，该攻击的日益完备给机器学习带来了严重的隐私威胁。本文从机器学习模型的攻防基础理论出发，分析成员推理攻击关键技术模型，厘清成员推理攻击模型与隐私泄露风险之间的关系，以期保证数据的隐私安全并促进机器学习应用领域的发展。首先，介绍了成员推理攻击的敌手模型、定义、分类以及攻击模型的生成机理。其次，分类总结和对比分析了成员推理攻击的攻击算法。然后，介绍了成员推理攻击在现实生活中的应用，并对成员推理攻击的防御技术进行了分类概括和对比分析。最后，通过对比分析已有的成员推理攻击方案及其防御技术方法，对机器学习成员推理攻击的发展趋势以及数据隐私保护的未来研究挑战进行展望。该工作为解决数据的隐私泄露问题提供一定的理论基础，对推动机器学习应用领域的发展有一定意义。

Abstract: Membership inference attacks can infer whether the target data is a member of a training dataset by attacking machine learning model, and the increasingly complete attack model poses a serious privacy threat to machine learning. Starting from the basic theory of attack and defense of machine learning models, this paper analyzed the key technical models and clarified the relationship between attack models and privacy leakage risks for ensuring the security of data privacy and promoting the development of machine learning applications field. Firstly, this paper introduced the adversary model of membership inference attacks, definition, classification and generation mechanism of the model. Secondly, we summarized and analyzed various existing membership inference attack algorithms. Then, the application of membership inference attack in real life was introduced, and the defense techniques of membership inference attack was classified and compared. Finally, by comparing and analyzing the existing attack schemes and their defense technology methods, the development trend of membership inference attack in machine learning and the future research challenges of data privacy protection are prospected. This work provides a theoretical basis for solving the problem of data privacy leakage, which is of great significance for promoting the development of machine learning applications.