Quantification of Data Uniqueness by Federated Learning with Autonomous Client Models

Abstract

Recently, Federated Learning (FL) has been attracting attention as a method for learning from distributed data. While FL has advantages such as privacy protection and reduced data traffic, it is difficult to characterize the non-i.i.d data of each client since data collection is not performed on the server side. In this study, we quantify the uniqueness of the data using FL with autonomous client models, which deals with a general model corresponding to all data constructed by the server side and a personalized model for each client data. The validity of the proposed method is assessed in terms of model performance and data feature extraction.