Machine Learning with Distributed Processing using Secure Divided Data: Towards Privacy-Preserving Advanced AI Processing in a Super-Smart Society

Abstract

Towards the realization of a super-smart society, AI analysis methods that preserve the privacy of big data in cyberspace are being developed. From the viewpoint of developing machine learning as a secure and safe AI analysis method for users, many studies have been conducted in this field on 1) secure multiparty computation (SMC), 2) quasi-homomorphic encryption, and 3) federated learning, among other techniques. Previous studies have shown that both security and utility are essential for machine learning using confidential data. However, there is a trade-off between these two properties, and there are no known methods that satisfy both simultaneously at a high level. In this paper, as a superior method in both privacy-preserving of data and utility, we propose a learning method based on distributed processing using simple, secure, divided data and parameters. In this method, individual data and parameters are divided into multiple pieces using random numbers in advance, and each piece is stored in each server. The learning of the proposed method is achieved by using these data and parameters as they are divided and by repeating partial computations on each server and integrated computations at the central server. The advantages of the proposed method are the preservation of data privacy by not restoring the data and parameters during learning; the improvement of usability by realizing a machine learning method based on distributed processing, as federated learning does; and almost no degradation in accuracy compared to conventional methods. Based on the proposed method, we propose backpropagation and neural gas (NG) algorithms as examples of supervised and unsupervised machine learning applications. Our numerical simulation shows that these algorithms can achieve accuracy comparable to conventional models.