Competitive and Asynchronous Decentralized Federated Learning with Blockchain Smart Contracts

Abstract

In recent years, machine learning models have evolved, and the training of these models requires large amounts of data. However, the training data often contains sensitive information, raising privacy concerns. Federated Learning has been proposed as a solution to mitigate privacy risks. Despite its advantages, Federated Learning still faces challenges such as the aggregator being a single point of failure, the existence of malicious participants, and the lack of incentives. Combining Federated Learning with blockchain technology could potentially address these challenges. In this study, we propose a new method for asynchronous Federated Learning using blockchain smart contracts. Our proposed method operates autonomously and in a decentralized manner without the need to trust any central organization, making it trustless. We propose an algorithm that motivates workers to submit high-quality models as quickly as possible. Workers’ behaviors are driven by incentive mechanisms. We deployed a smart contract on a local Ethereum blockchain and executed multiple workers. Our evaluation results demonstrate that learning converges and achieves accuracy comparable to conventional Federated Learning, indicating the effectiveness of our proposed method.