A Decentralized AI Data Management System In Federated Learning

Federated Learning is a distributed machine learning approach which enables model training without sharing private locally produced data. It has been actively researched for several years as a means to utilize big data while protecting personal information. However, the server must decide which clients to participate in and what results to be used for aggregation each round. Besides, since the server needs to maintain the connection with the client directly, device overload and the processing delay may cause due to changes in the system environment such as network condition. In this paper, we propose a data management system that efficiently addresses the problem of general Federated Learning by improvements of the data management process on the connection between the Federated Learning server and the client. Additionally, it is shown that the proposed system can perform tasks independently and scales for increasing number of devices participating in the Federated Learning tasks.