A Method of Federated Learning Based on Blockchain

Abstract

Currently many enterprises face issues regarding insufficient data collection samples and data recording dimensions, thus it's hard to make efficient predictions. Since it is limited by the requirement of protecting privacy and trade secrets, data can't be effectively shared among enterprises. Federated learning is an effective method to solve this problem, but there are some performance bottlenecks, information security issues and data trust issues still existed, which need to be improved in combination with other advanced technologies to meet the practical requirements. This paper combines the blockchain technology with federated learning technology, and uses decentralized blockchain system to replace the traditional centralized federated learning architecture. We adopt training method of updating models to achieve machine learning. In this way, we can avoid transmission of intermediate computing data and achieve mechanism of node access, model evaluation, motivation and audit with combination of block chain. In terms of the algorithm, the horizontal federated learning adopts the integrated learning algorithm, and the vertical federated learning adopts the deep learning algorithm. It will be described in detail below.