Scholarlink

<p>Edge Federated Learning (EFL), formed by integrating edge computing and federated learning, faces constraints from resource heterogeneity and unequal benefit distribution. This study constructs a dynamic incentive mechanism optimization framework based on Stackelberg game theory. The framework regulates edge node behavior through a leader–follower game model, dynamic reward and punishment strategies, and an improved Game-based Incentive Mechanism Algorithm (GIMA) (defined as a game-theoretic iterative optimization algorithm for incentive strategy adaptation in heterogeneous edge environments). Simulations using classical datasets including MNIST, CIFAR-10, IMDB, and Human Activity Recognition (HAR) verify its effectiveness. In the scenario with 100 heterogeneous edge nodes, the node participation rate reaches 89.3% ± 2.3, the number of model convergence rounds decreases by 40.5% ± 3.1, the incentive cost reduces by 25.8% ± 2.7, and the test accuracy remains at 91.8% ± 0.7. Even under dynamic node exit with a 20% exit rate or extreme heterogeneity with a variance of 0.5, the mechanism maintains over 79% node participation. It also balances privacy protection and system efficiency, achieving 82.1% participation when the differential privacy budget ε is 0.1. Under the experimental conditions of this study, the proposed mechanism helps improve node participation, shorten training cycles, and reduce resource consumption for EFL systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The optimization of incentive mechanisms for edge federated learning based on Stackelberg game

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Liping He

摘要

Edge Federated Learning (EFL), formed by integrating edge computing and federated learning, faces constraints from resource heterogeneity and unequal benefit distribution. This study constructs a dynamic incentive mechanism optimization framework based on Stackelberg game theory. The framework regulates edge node behavior through a leader–follower game model, dynamic reward and punishment strategies, and an improved Game-based Incentive Mechanism Algorithm (GIMA) (defined as a game-theoretic iterative optimization algorithm for incentive strategy adaptation in heterogeneous edge environments). Simulations using classical datasets including MNIST, CIFAR-10, IMDB, and Human Activity Recognition (HAR) verify its effectiveness. In the scenario with 100 heterogeneous edge nodes, the node participation rate reaches 89.3% ± 2.3, the number of model convergence rounds decreases by 40.5% ± 3.1, the incentive cost reduces by 25.8% ± 2.7, and the test accuracy remains at 91.8% ± 0.7. Even under dynamic node exit with a 20% exit rate or extreme heterogeneity with a variance of 0.5, the mechanism maintains over 79% node participation. It also balances privacy protection and system efficiency, achieving 82.1% participation when the differential privacy budget ε is 0.1. Under the experimental conditions of this study, the proposed mechanism helps improve node participation, shorten training cycles, and reduce resource consumption for EFL systems.