Federated Graph Neural Networks: Overview, Techniques, and Challenges

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Graph Neural Networks (GNNs) are powerful for graph data but face privacy challenges.
• Federated Learning (FL) offers privacy by training on decentralized data.
• Client relationships in FL present opportunities for performance enhancement.

Purpose of the Study:

• To provide a comprehensive survey of the emerging field of Federated Graph Neural Networks (FedGNNs).
• To offer a structured understanding of how GNNs and FL are integrated and how FedGNNs handle client heterogeneity.
• To identify key challenges and future research directions in FedGNNs.

Main Methods:

• A comprehensive literature review of FedGNNs.
• Proposal of a 2-D taxonomy: main taxonomy (GNN-FL integration) and auxiliary taxonomy (heterogeneity).
• Discussion of existing works, challenges, and limitations.

Main Results:

• A structured overview of FedGNN research, categorizing approaches based on GNN-FL integration and heterogeneity handling.
• Identification of critical challenges including robustness, explainability, efficiency, fairness, and inductive capabilities.
• Envisioning future research avenues for advancing FedGNN technology.

Conclusions:

• FedGNNs represent a rapidly developing interdisciplinary field with significant potential.
• A clear taxonomy and discussion of challenges are crucial for researchers entering this field.
• Future work should focus on developing more robust, explainable, efficient, fair, inductive, and comprehensive FedGNNs.