Revolutionizing machine learning: Blockchain-based crowdsourcing for transparent and fair labeled datasets supply

Abstract

With the development of 5G/6G communication networks, the industrial Internet of Things (IIoT) industry has generated a massive amount of data, presenting opportunities for advancements in the field of machine learning. The core of machine learning, labeled datasets, requires qualities such as diversity, quantity, and quality. However, the current collection of training datasets is mostly centralized, relying on crowdsourcing systems as platforms. Hence, there are four open challenges existing: (1) traditional crowdsourcing systems are mostly based on centralized platform which often suffers single point of failure, mischief attacks, DDoS attacks and are easy to be remotely hijacked; (2) the transparency of traditional crowdsourcing systems cannot be guaranteed, which results in the unfairness in assigning task or rewarding worker; (3) the quality of labeled training dataset cannot be guaranteed as the workers in traditional crowdsourcing systems are professional or non-professional; (4) the privacy of labels cannot be preserved as the platform can also learn the labeled result. In this paper, we utilize web 3.0 technology to propose a blockchain-based labeled training dataset supply system which can simultaneously supply annotation service and labeled training dataset with the challenges above overcome. Meanwhile we design a privacy-preserved truth discovery suitable for categorical data by combining it with Software Guard Extensions (SGX). Furthermore, we design a fair rewards distribution mechanism which is based on reputation system and Shapley value. The two mechanisms above can ensure the quality of labeled training dataset. Finally, to demonstrate the practicability of our design, we implement a prototype of the system deployed on Fabric test network and conduct extensive simulations. Compared to applications on public chains, the throughput of our blockchain application can reach 53 times higher.