BlockFed: A Novel Federated Learning Framework Based On Hierarchical Aggregation

Abstract

This paper presents a pioneering federated learning framework that leverages a novel hierarchical aggregation approach, empowering clients to collaboratively generate the global model through multiple levels of aggregation. Additionally, a unique role definition mechanism is integrated into the framework to delineate clients’ roles and tasks in each learning round. Moreover, decentralized storage (e.g. IPFS) and blockchain technologies are employed for storing local models and their corresponding hash pointers, respectively, to improve data availability and integrity. We implemented our solution using using Keras, Scikit-learn, web3 and Solidity. The performance of the proposed framework is evaluated using a genomic breast cancer dataset sourced from the GDC portal, yielding a remarkable 98% accuracy for the global model after 12 rounds of learning while for a centrally trained model on the 70% of the entire dataset the accuracy was 95%. This clearly shows the effectiveness of the proposed framework.