Simple Privacy-Preserving Federated Learning with Different Encryption Keys

Federated learning is a method where multiple participants collaboratively train a model using machine learning techniques, such as deep learning, while keeping each participant's data private through the use of a central server. However, there are cases in which participant input information is leaked to the server. To solve this problem, several protocols have been proposed. In particular, Phong et al. (in IEEE TIFS 2018) proposed a protocol that protects participant information against a server using homomorphic encryption. Later, Park et al. (in ICTC 2022) proposed an improved protocol, where each client has a different secret-key for the homomorphic encryption. In this paper, we show that the Park et al.'s protocol has some weakness and propose the secure one based on the previously proposed protocols.

I. Introduction

Background: Machine learning has greatly contributed to the development of society in recent years playing a pivotal role in several important fields, such as finance and agriculture. However, some issues need to be resolved when using machine learning. For example, how to handle client privacy when the dataset used for learning contains personal information. If a single organization holds a large amount of personal information, this is not a major problem because clients can learn their model without the help of other organizations. However, when multiple organizations hold personal information, while aggregating the information for learning, then the issue of privacy arises.