Split learning is a novel technology developed by MIT Media lab’s Camera culture group that allows to train ML models without sharing any raw data and overcomes the drawbacks of Federated learning.
Published in Chapter:
A Blockchain-Based Federated Learning: Concepts and Applications
Ankit Khushal Barai (Department of CSE, Indian Institute of Information Technology, Nagpur, India), Robin Singh Bhadoria (Department of Computer Science and Engineering, Hindustan College of Science and Technology, India), Jyotshana Bagwari (Department of CSE, Uttarakhand Technical University, India), and Ivan A. Perl (ITMO University, Russia)
Copyright: © 2021
|Pages: 20
DOI: 10.4018/978-1-7998-5876-8.ch008
Abstract
Conventional machine learning (ML) needs centralized training data to be present on a given machine or datacenter. The healthcare, finance, and other institutions where data sharing is prohibited require an approach for training ML models in secured architecture. Recently, techniques such as federated learning (FL), MIT Media Lab's Split Neural networks, blockchain, aim to address privacy and regulation of data. However, there are difference between the design principles of FL and the requirements of Institutions like healthcare, finance, etc., which needs blockchain-orchestrated FL having the following features: clients with their local data can define access policies to their data and define how updated weights are to be encrypted between the workers and the aggregator using blockchain technology and also prepares audit trail logs undertaken within network and it keeps actual list of participants hidden. This is expected to remove barriers in a range of sectors including healthcare, finance, security, logistics, governance, operations, and manufacturing.