Federated Node Classification over Distributed Ego-Networks with Secure Contrastive Embedding Sharing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Federated graph learning (FGL) enables collaborative learning on distributed graph data while preserving data privacy.
• Existing FGL methods often assume clients hold entire graphs or graph partitions, not individual ego-networks.
• Ego-network settings present unique challenges due to incomplete neighborhood information for non-ego nodes.

Purpose of the Study:

• To propose a novel FGL method tailored for distributed ego-networks.
• To address the challenge of incomplete neighborhood information in ego-network settings.
• To enhance privacy and accuracy in federated learning on decentralized graph data.

Main Methods:

• Developed an FGL approach for distributed ego-networks involving node embedding sharing.
• Implemented a contrastive learning mechanism to align local and global node embeddings.
• Utilized a secure embedding sharing protocol to protect node identity and privacy.

Main Results:

• Demonstrated the effectiveness of the proposed embedding sharing method across various federated model sharing frameworks.
• Showcased improved performance in FGL tasks utilizing ego-network data.
• Validated the method's ability to handle incomplete neighborhood information.

Conclusions:

• The proposed FGL method effectively addresses challenges in distributed ego-networks.
• Embedding sharing with contrastive learning and secure protocols enhances privacy and model performance.
• Further research is needed to mitigate potential efficiency and privacy drawbacks.