Federated Data Integration for Heterogeneous Partitions Based on Differential Privacy

Abstract

Federated learning has recently become a research hotspot in distributed learning, and its purpose is to jointly train machine learning models on the premise of protecting privacy. However, there are some problems with federated learning. Each machine learning algorithm must be modified in order to complete the training. The data partitioning is either horizontal or vertical, which is not flexible enough. In addition, there are many rounds of communication during the training process, so the training efficiency is low. In order to address these problems, we propose a generic federated integration method for multiple data sources. The method can integrate data in arbitrary partitions, protect the privacy based on differential privacy, and reduce communication cost based on singular value decomposition. After the data are modelled in this method, they can be transferred to the center for purpose of federated learning. Our method includes four algorithms. We give a theoretical proof on the method’s satisfying of differential privacy. Finally, experiments are conducted to demonstrate the performance of the method in prediction accuracy and data compression.