Blockchain-Based Decentralized Federated Learning

Abstract

Federated learning (FL) has gained great traction in recent years. It can provide a privacy-preserving mechanism to train machine learning models on hidden data. However, most of today's FL systems use a centralized server to build the global FL model. Such centralization raises trust and fairness issues stemming from the fact that the FL server may have the ability to reconstruct the original data successfully. In this paper, we propose a blockchain-based decentralized FL system. The FL process in the proposed system is composed of two stages. In the first stage, FL nodes reach a consensus on the training configurations using smart contracts. In the second stage, nodes aggregate the model updates using a novel decentralized aggregation method. The proposed system efficiently schedules aggregation tasks between decentralized nodes, handles dropouts, and detects malicious acts. We test and analyze our solution using Ethereum smart contracts and Python. The implementation and testing details of the algorithms are presented, and all codes are publicly available on GitHub.