Scholarlink

<p>Efficient data utilization and strong privacy protection are major challenges in Intelligent Transportation Systems (ITS), particularly in complex environments with highly distributed Intelligent Connected Vehicles (ICVs). Conventional machine learning methods struggle to capture complex spatiotemporal dependencies while maintaining data privacy and locality. To overcome these limitations, we propose FedGDAN, a Federated Graph Diffusion Attention Network that combines graph neural networks (GNNs) with federated learning (FL) to enable collaborative traffic flow prediction without sharing raw data. FedGDAN models global spatiotemporal correlations across road networks and introduces an adaptive local aggregation mechanism to address non-independent and identically data distributions, thereby improving robustness and accuracy. Experiments on real-world datasets show that FedGDAN consistently outperforms state-of-the-art centralized and federated baselines, achieving 3%–10% gains in Mean Absolute Error.</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

FedGDAN: Privacy-preserving traffic flow prediction via federated graph diffusion attention networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月20日

查看全文
  • Yuanhui Li,
  • Bo Mi,
  • Ran Zeng

摘要

Efficient data utilization and strong privacy protection are major challenges in Intelligent Transportation Systems (ITS), particularly in complex environments with highly distributed Intelligent Connected Vehicles (ICVs). Conventional machine learning methods struggle to capture complex spatiotemporal dependencies while maintaining data privacy and locality. To overcome these limitations, we propose FedGDAN, a Federated Graph Diffusion Attention Network that combines graph neural networks (GNNs) with federated learning (FL) to enable collaborative traffic flow prediction without sharing raw data. FedGDAN models global spatiotemporal correlations across road networks and introduces an adaptive local aggregation mechanism to address non-independent and identically data distributions, thereby improving robustness and accuracy. Experiments on real-world datasets show that FedGDAN consistently outperforms state-of-the-art centralized and federated baselines, achieving 3%–10% gains in Mean Absolute Error.