首页|期刊导航|网络与信息安全学报|跨客户端邻居聚合与注意力融合的联邦图神经网络框架

跨客户端邻居聚合与注意力融合的联邦图神经网络框架OA

Federated graph neural network framework for cross-client neighbor aggregation and attention fusion

中文摘要英文摘要

联邦图神经网络将图神经网络与联邦学习相结合,使各方能够协作训练模型而无须共享私有图数据.但训练阶段图数据被划分并分配至多台客户端,跨客户端的边导致节点邻居信息可能存储在不同设备上,从而引发信息丢失.同时,各客户端持有的子图通常来自全局图的不同部分(如不同社区),导致数据分布存在显著异质性.针对分布式图数据中跨客户端边信息缺失与数据异质性这两大挑战,提出了一种联邦图卷积神经网络训练框架.该框架在预训练阶段采用基于同态加密的跨客户端邻居特征聚合方法,以安全的方式弥补局部结构信息缺口;在正式训练阶段设计了一种基于随机图嵌入的注意力联邦聚合机制,通过动态评估客户端之间的相似性实现智能权重分配,避免异质数据的强行融合.同时,结合自适应稀疏掩码技术,实现参数层面的个性化筛选与高效更新.在Cora、CiteSeer和PubMed数据集上的实验结果表明,所提框架在节点分类任务上的准确率相比主流基线框架最高提升了11.58%.

Federated graph neural networks combine graph machine learning with federated learning,allowing par-ties to collaborate on training models without sharing private graph data.However,due to the partitioning of graph data into multiple clients during training,the existence of cross-client edges may result in node neighbor information being stored on different devices,leading to information loss.Meanwhile,the subgraphs held by each client often come from different parts of the global graph(such as different communities),resulting in significant heterogeneity in data distribution.To address the two major challenges of missing cross-client edge information and data heteroge-neity in distributed graph data,a federated graph convolutional neural network training framework was proposed.This framework used homomorphic encryption based secure cross-client neighbor feature aggregation in the pre-training round to compensate for local structural information gaps;An attention federated aggregation mechanism based on random graph embedding was designed and used in the training round,which dynamically evaluates client similarity to achieve intelligent weight allocation and avoid forced fusion of heterogeneous data.At the same time,by integrating adaptive sparse masking technology,personalized filtering and efficient updating are achieved at the pa-rameter level.The experimental results on the Cora,CiteSeer and PubMed datasets show that the proposed framework achieves an accuracy improvement of up to 11.58%over the mainstream baselines in the node classification task.

邓赫;宋秀丽

重庆邮电大学网络空间安全与信息法学院,重庆 400065重庆邮电大学网络空间安全与信息法学院,重庆 400065

信息技术与安全科学

联邦学习图神经网络联邦图神经网络同态加密

federated learninggraph neural networkfederated graph neural networkhomomorphic encryption

《网络与信息安全学报》 2026 (4)

134-147,14

10.11959/j.issn.2096-109x.AQ26048

评论