Subgraph-level federated graph neural network for privacy-preserving recommendation with meta-learning

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Privacy

Background:

• Traditional graph neural networks (GNNs) in recommendation systems face privacy challenges due to centralized data handling.
• Federated learning offers a decentralized approach but requires robust privacy-preserving mechanisms.

Purpose of the Study:

• To develop a federated framework for privacy-preserving GNN-based recommendations.
• To enhance data privacy and model performance in distributed recommendation systems.

Main Methods:

• Implemented a federated framework for distributed GNN training using local user data.
• Utilized Software Guard Extension (SGX) for secure subgraph exchange and expansion.
• Applied Local Differential Privacy (LDP) to ensure data privacy during gradient aggregation.
• Incorporated Prototype Networks (PN) and Model-Agnostic Meta-Learning (MAML) for personalized recommendations and handling data heterogeneity.

Main Results:

• The proposed federated framework significantly outperforms centralized GNN-based recommendation methods.
• The system effectively preserves user privacy while maintaining high recommendation accuracy.
• Demonstrated superiority across six diverse datasets.

Conclusions:

• The federated GNN framework offers a viable solution for privacy-preserving recommendations.
• SGX and LDP effectively mitigate privacy risks in federated learning environments.
• Personalization techniques improve adaptability to heterogeneous client data.