Proceedings of the 3rd International Conference on Big Data Economy and Digital Management, BDEDM 2024, January 12–14, 2024, Ningbo, China

Research Article

Incentive Strategies of Clients in Decentralized Federated Learning Using Evolutionary Game

Download83 downloads
  • @INPROCEEDINGS{10.4108/eai.12-1-2024.2347204,
        author={Qingjing  Feng and Yaqian  Ma},
        title={Incentive Strategies of Clients in Decentralized Federated Learning Using Evolutionary Game},
        proceedings={Proceedings of the 3rd International Conference on Big Data Economy and Digital Management, BDEDM 2024, January 12--14, 2024, Ningbo, China},
        publisher={EAI},
        proceedings_a={BDEDM},
        year={2024},
        month={6},
        keywords={blockchain decentralized federated learning evolutionary game evolutionarily stable strategy (ess) incentive mechanism},
        doi={10.4108/eai.12-1-2024.2347204}
    }
    
  • Qingjing Feng
    Yaqian Ma
    Year: 2024
    Incentive Strategies of Clients in Decentralized Federated Learning Using Evolutionary Game
    BDEDM
    EAI
    DOI: 10.4108/eai.12-1-2024.2347204
Qingjing Feng1,*, Yaqian Ma1
  • 1: Beijing University of Technology
*Contact email: fengqingjing@emails.bjut.edu.cn

Abstract

The decentralized federated learning framework combined with blockchain technology can significantly enhance the data credibility and security of traditional federated learning. To solve challenges like low data quality and "free riding", it is essential to introduce more effective incentive strategies that can be practically applied. In this paper, we propose a more general decentralized federated learning framework to establish an evolutionary game model with twelve variables that affect the incentive strategies of clients. By discussing the interaction mechanism among clients, we aim to determine effective incentive strategies. Additionally, we conduct numerical simulation to analyze the impact of these twelve variables on stable strategies in the evolutionary game. The results demonstrate the effectiveness of our proposed model and associated algorithms. The experiment results indicate that an increase in the willingness to cooperate of either party will increase the probability of the overall evolution towards the optimal outcome. Factors such as reducing training costs, adjusting incentive levels, minimizing additional losses, increasing the recognition probability, and offering additional benefits can expedite the implementation of the evolutionarily stable strategy (ESS) among clients.This study can provide valuable insights to improve the training quality of clients and boost overall model efficiency.