Dynamic Incentive Design for Federated Learning Based on Consortium Blockchain Using a Stackelberg Game

Abstract

Federated learning (FL) has emerged as a powerful privacy-preserving approach that enables multiple devices to collaboratively train a machine learning model without sharing raw data. To motivate client participation and protect against malicious threats such as poisoning attacks, it is necessary to design a fair trading platform with an incentive mechanism. In this paper, we propose a blockchain-based FL framework and design a dynamic incentive mechanism to establish a decentralized and transparent trading platform. A reputation mechanism is employed to assess the reliability of participants, and blockchain technology is utilized to ensure secure and tamper-proof management. Additionally, we design a dynamic incentive mechanism based on reputation and Stackelberg game theory, where the task publisher, acting as the leader, determines the total reward, and the participants, as followers, decide their level of engagement. Participants are selected and rewarded based on their reputations and bids, within a constrained budget. Experimental results on MNIST, Fashion-MNIST and CIFAR-10 datasets demonstrate that our proposed framework effectively promotes high-quality participants to engage in the training task. It prevents malicious participants from disrupting the training process, maintaining the high accuracy across different scenarios, even with 50% malicious participants.