Machine learning for privacy-preserving: Approaches, challenges and discussion

Abstract

Currently, advanced technologies such as big data, artificial intelligence and machine learning are undergoing rapid development. However, the emergence of cybersecurity and privacy leakage problems has resulted in serious implications. This paper discusses the current state of privacy security issues in the field of machine learning in a comprehensive manner. During machine training, training models often unconsciously extract and record private information from raw data, and in addition, third-party attackers are interested in maliciously extracting private information from raw data. This paper first provides a quick introduction to the validation criterion in privacy-preserving strategies, based on which algorithms can account for and validate the privacy leakage problem during machine learning. The paper then describes different privacy-preserving strategies based mainly on federation learning that focus on Differentially Private Federated Averaging and Privacy-Preserving Asynchronous Federated Learning Mechanism and provides an analysis and discussion of their advantages and disadvantages. By improving the original machine learning methods, such as improving the parameter values and limiting the range of features, the possibility of privacy leakage during machine learning is successfully reduced. However, the different privacy-preserving strategies are mainly limited to changing the parameters of the original model training method, which leads to limitations in the training method, such as reduced efficiency or difficulty in training under certain conditions.