Blockchain-enabled federated learning with mechanism design

Abstract

Federated learning (FL) is a promising decentralized deep learning technique that allows users to collaboratively update models without sharing their own data. However, due to its decentralized nature, no one can monitor workers’ behavior, and they may thus deviate protocols (e.g., participating without updating any models). To solve this problem, many researchers have proposed blockchain-enabled FL to reward workers (or users) with cryptocurrencies to encourage workers to follow the protocols. However, there is a lack of theoretical discussions concerning how such rewards impact workers’ behavior and how much should be given to workers. In this paper, we propose a mechanism-design-oriented FL protocol on a public blockchain network. Mechanism design (MD) is often used to make a rule intended to achieve a specific goal. With MD in mind, we introduce the concept of competition into blockchain-based FL so that only workers who have contributed well can obtain rewards, which naturally prevents workers from deviating from the protocol. We then mathematically answer the following questions with contest theory, a novel field of study in economics: i) What behavior will workers take?; ii) how much effort should workers exert to maximize their profits?; iii) how many workers should be rewarded?; and iv) what is the best proportion for reward distribution?.