The Design of Reputation System for Blockchain-based Federated Learning

Federated learning (FL) is a new solution to fulfill machine learning (ML) in a decentralized manner. In FL, a group of participants train their local models using their private data, and then upload the locally trained models to a central server that completes the model aggregation process. Compared with the traditional ML that needs to upload private data of the participants to a server for a centralized training, FL protects the data privacy of participants. However, the existing FL has several drawbacks: FL relies to a large extent on a single central server that takes the risk of single point of failure; it lacks a fair and effective reward system to prevent malicious participants from damage the learning process. To overcome these drawbacks, this paper proposes a new FL system based on blockchain (BC). With BC, we introduce a voting-based reputation mechanism to guarantee that the locally trained models from effective FL participants (other than malicious FL participants) can be selected to improve the efficiency of model aggregation. Besides that, we combine the reputation system with a reward mechanism to dynamically adjust the reward shares that can be fairly allocated to the participants. This can attract more FL participants to participate in the learning process, so that the system can run stably for a long time. The experimental results show that our system can achieve more efficient FL learning and can also defend against malicious attacks to a certain extent.

I. Introduction

Federated learning (FL) was first proposed by Google as a new Machine Learning (ML) framework [1]–[3]. With the framework of FL, android mobile phone terminals can train their local models using their own data on the phones, and then send the trained local models to a cloud server that aggregates all local models into a global model. In the learning process of FL, the cloud server does not know each terminal's data, but its aggregated model can benefit from all data at the terminals [4]. This FL approach solves the problem of data leakage in traditional ML by allowing data owners to store data in their own hands but training a globe model using all data [2].