A federated graph neural network framework for privacy-preserving personalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Graph Neural Networks (GNNs) excel at modeling complex interactions for personalized applications like recommendations.
• Current GNN personalization methods use centralized learning, posing significant privacy risks due to sensitive user data.
• Decentralized data necessitates novel approaches for effective and privacy-preserving GNN training.

Purpose of the Study:

• To introduce FedPerGNN, a federated GNN framework for effective and privacy-preserving personalization.
• To enable collaborative training of GNN models on decentralized graphs while safeguarding user privacy.
• To enhance personalization by incorporating higher-order graph information through a privacy-preserving expansion protocol.

Main Methods:

• Developed a privacy-preserving model update mechanism for collaborative GNN training on local, decentralized graphs.
• Introduced a privacy-preserving graph expansion protocol to integrate distant graph information without compromising data security.
• Evaluated FedPerGNN on six diverse datasets across various personalization scenarios.

Main Results:

• FedPerGNN demonstrated superior performance, achieving 4.0%–9.6% lower errors compared to state-of-the-art federated personalization methods.
• The framework successfully maintained strong privacy protection throughout the decentralized training process.
• Experimental results validate the effectiveness of FedPerGNN in diverse personalization tasks.

Conclusions:

• FedPerGNN offers a robust solution for privacy-preserving personalization using decentralized graph data.
• The framework effectively leverages GNNs in a federated setting, mitigating privacy concerns associated with centralized approaches.
• FedPerGNN represents a significant advancement for responsible and intelligent personalization by mining decentralized graph information.