AUTHOR=Liu Yanjun , Li Hongwei , Hao Meng 

TITLE=Personalized and privacy-preserving federated graph neural network

JOURNAL=Frontiers in Physics

VOLUME=Volume 12 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2024.1383276

DOI=10.3389/fphy.2024.1383276

ISSN=2296-424X

ABSTRACT=Cyber-physical-social systems (CPSSs) seamlessly connects networks, physical devices and social spaces through data. CPSS provides a more comprehensive intelligent system for federated graph neural networks, thus promoting the rapid development of artificial intelligence(AI). Graph neural networks(GNN) have emerged as one of the most critical research areas in general AI.High-performance GNN obtains dependencies within a graph by capturing the mechanism of message passing and aggregation between neighboring nodes in the graph, and successfully updates node embeddings. However, in practical applications, the inherent model structure of the graph is highly susceptible to privacy attacks, and the heterogeneity of external data can lead to a decrease in model performance. Motivated by this challenge, this work proposes a novel framework called Personalized Federated Graph Neural Network for Privacy-Preserving (PFGNN), by which a high-quality and highly secure global model is trained collaboratively by multiple clients. Specifically, firstly, this work introduces a graph similarity strategy. Based on the principle that clients with similar features exhibit stronger homophily, this work divides all participating clients into multiple clusters for collaborative training. Furthermore, within each group, this work employs an attention mechanism to design a federated aggregation weighting scheme. This scheme is used to construct a global model on the server, which helps mitigate the difficulty of model generalization resulting from data heterogeneity collected from different clients. Lastly, to ensure the privacy of model parameters during the training process and prevent malicious adversaries from stealing them, this work implements privacy-enhancing technology by introducing an optimized functionhiding multi-input function encryption scheme. This ensures the security of both model data and user privacy. In comparison to state-of-the-art works, our communication overhead scales linearly with the number of clients. Our work provides the foundation for designing efficient and privacy-protecting personalized federated graph neural networks.