Privacy-enhanced Graph Edge Convolutional Networks

Graphs are a fundamental structure in nature and data, representing relationships across domains such as biology, social networks, and neuroscience. With the rise of Graph Neural Networks (GNNs), researchers have harnessed the inherent properties of graphs to solve complex problems. Among these, Graph Edge Convolution Networks (GECNs) have emerged as a versatile variant of GNNs, uniquely focusing on the edges - relationships between nodes - rather than just the nodes themselves. This makes GEC particularly powerful for applications where the nature of connections holds critical information. Such modern model deployment in real-world scenarios often necessitates running on external servers, posing significant risks to the privacy of sensitive data. Fully Homomorphic Encryption (FHE) offers a promising solution by enabling computations on encrypted data without ever needing to decrypt it, safeguarding privacy throughout the model's operation. By implementing GEC layers in FHE using the HEAAN library, we show that input data remains fully encrypted during the forward pass, ensuring maximum data protection. In this paper, we demonstrate how our FHEbased GEC model maintains accuracy while processing encrypted data, validated through a sample application on brain activity analysis. Leveraging advanced FHE features like SIMD, we overcome encryption limitations and demonstrate the practicality of privacy-enhanced GNNs for sensitive applications.