Scholarlink

Graph Convolutional Networks (GCN) are widely used for graph-based learning but face challenges in large-scale, privacy-preserving, and distributed settings. FedAKD, a novel federated learning framework, addresses issues such as data heterogeneity, inter-client communication costs, and varying client computing power. It clusters clients based on functional similarity of GCN outputs, aggregates node information within clusters, and uses Paillier homomorphic encryption for secure data enhancement. The top-ranked client receives the enhanced data, and knowledge distillation is employed to create efficient student models for low-power clients, reducing communication overhead. Experiments on Cora, CiteSeer, and Pubmed datasets show FedAKD’s superior performance and efficiency compared to baseline methods.

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

FedAKD: Heterogeneous Graph Federated Learning Framework Based on Data Augmentation and Knowledge Distillation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tengfei Cao,
  • Yiming Zhang

摘要

Graph Convolutional Networks (GCN) are widely used for graph-based learning but face challenges in large-scale, privacy-preserving, and distributed settings. FedAKD, a novel federated learning framework, addresses issues such as data heterogeneity, inter-client communication costs, and varying client computing power. It clusters clients based on functional similarity of GCN outputs, aggregates node information within clusters, and uses Paillier homomorphic encryption for secure data enhancement. The top-ranked client receives the enhanced data, and knowledge distillation is employed to create efficient student models for low-power clients, reducing communication overhead. Experiments on Cora, CiteSeer, and Pubmed datasets show FedAKD’s superior performance and efficiency compared to baseline methods.