Federated Machine Learning using Blockchain Technology

Abstract

Recently, Federated Learning (FL) has gained tremendous traction as it has the ability to provide a privacy-preserving mechanism to train Machine Learning models on hidden data. However, most of today's FL systems are centralized, in which a centralized server is typically used to build the global FL model. Such centralization raises trust and fairness issues stemmed from the fact that the FL server may have the ability to reconstruct the original data successfully. To address this, we propose a fair blockchain-based decentralized FL system. In this thesis, we start by building a secure decentralized FL system from scratch in chapter 3, we integrate it with blockchain to achieve full decentralization in chapter 4. Finally, we present and test our own scheme to ensure fairness in chapter 5. In the proposed system, the model aggregation is performed with no server involvement, therefore, we introduce a novel decentralized aggregation method to schedule and manage the aggregation process. This method has comparable performance to the centralized approach, it handles nodes dropping out in an efficient and deterministic manner, and the aggregation work of any node can be verified. We test all components in the proposed system and analyze the security, performance, and cost aspects.

Date of Award	Dec 2021
Original language	American English