A Privacy Preserving Scheme with Dimensionality Reduction for Distributed Machine Learning

Abstract

To obtain useful results in machine learning, it is required to collect data from multiple institutions and learn with larger-scale data. However, data collected from multiple institutions may contain a lot of personal information and should not be explicitly shared. The existing research has proposed various methods to protect privacy by using encryption or anonymization, but encryption causes large computational costs, and anonymization may greatly reduce the usefulness of data. In this research, we propose a privacy protection method using dimensionality reduction that is difficult to reverse while maintaining the high usefulness of data. The main idea of our method is that combining dimensionality reduction algorithms with noise addition is useful for privacy-preserving data analysis with high accuracy. Furthermore, we evaluate the effectiveness and security of this method and show the utility of the proposed method.