Scholarlink

<p>In the era of big data, knowledge learning fueled by massive datasets has paved the way for the development of a vast array of intelligent applications. Among these, Federated Learning (FL) enables collaborative model training without sharing raw data, but faces critical challenges when users’ gradients are heterogeneous (noisy, incomplete, or distributionally skewed). Such heterogeneity biases aggregation results and degrades model generalizability, while existing solutions overlook incentive-cost trade-offs. We propose a contract-based FL framework that addresses both technical and economic inefficiencies through two key innovations: (1) Personalized aggregation weights and training batches assigned via gradient quality metrics, and (2) A gradient-based framework for optimizing contract parameters, dynamically tailoring computational rewards to user-specific data distributions. These mechanisms ensure incentive compatibility while reducing budget waste. Experimental evaluations demonstrate that under the same cost constraints, the contract-based aggregation and discrimination training epoch can enhance the generalization accuracy of FL tasks,with contracts being incentive-compatible and flexible.</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

Optimize the model performance in federated learning: a contract theory-based heuristic approach

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Tao Wan,
  • Tiantian Jiang,
  • Weichuan Liao,
  • Nan Jiang

摘要

In the era of big data, knowledge learning fueled by massive datasets has paved the way for the development of a vast array of intelligent applications. Among these, Federated Learning (FL) enables collaborative model training without sharing raw data, but faces critical challenges when users’ gradients are heterogeneous (noisy, incomplete, or distributionally skewed). Such heterogeneity biases aggregation results and degrades model generalizability, while existing solutions overlook incentive-cost trade-offs. We propose a contract-based FL framework that addresses both technical and economic inefficiencies through two key innovations: (1) Personalized aggregation weights and training batches assigned via gradient quality metrics, and (2) A gradient-based framework for optimizing contract parameters, dynamically tailoring computational rewards to user-specific data distributions. These mechanisms ensure incentive compatibility while reducing budget waste. Experimental evaluations demonstrate that under the same cost constraints, the contract-based aggregation and discrimination training epoch can enhance the generalization accuracy of FL tasks,with contracts being incentive-compatible and flexible.