ESA-FedGNN: Efficient secure aggregation for federated graph neural networks.

Area of Science:

• Machine Learning
• Data Privacy
• Network Analysis

Background:

• Graph Neural Networks (GNNs) excel at node embedding but face privacy challenges due to centralized data storage.
• Federated Learning (FL) offers a privacy-preserving alternative, but secure aggregation in GNNs remains complex.
• Existing methods struggle with communication costs and computational redundancy in privacy-preserving GNNs.

Purpose of the Study:

• To propose an Efficient Secure Aggregation for Federated Graph Neural Network (ESA-FedGNN) to address privacy and efficiency concerns.
• To reduce communication overhead and computational redundancy in federated graph neural network training.
• To enhance data privacy and security against malicious attacks in distributed graph learning.

Main Methods:

• Developed a novel secret sharing scheme using Fast Fourier Transform (FFT) and Newton Interpolation for secure data sharing and reconstruction.
• Introduced regular graph embedding based on geometric distribution to optimize aggregation speed via data parallelism.
• Implemented a double mask strategy to safeguard model parameters from unauthorized access.

Main Results:

• Achieved significant improvements in efficiency and privacy compared to existing state-of-the-art methods (specific metrics omitted for brevity).
• Demonstrated reduced communication costs and computational redundancy in the federated learning process.
• Successfully prevented malicious adversaries from stealing model parameters, ensuring robust privacy.

Conclusions:

• ESA-FedGNN provides an efficient and secure solution for privacy-preserving federated graph neural networks.
• The proposed methods enhance the practical applicability of GNNs in sensitive domains like drug discovery and e-commerce.
• This research contributes to the development of secure and privacy-preserving machine learning technologies.