Privacy-Preserving Graph Neural Network Inference Using Homomorphic Encryption for Secure Graph Classification

Abstract

Graph Neural Networks have achieved remarkable success in graph-based learning tasks, but often require access to sensitive data during inference, raising privacy concerns in domains such as healthcare, finance, and cybersecurity. Homomorphic Encryption enables computation on encrypted data; however, directly applying HE to graph neural network operations remains computationally expensive due to complex message-passing and attention mechanisms. This paper presents Hybrid-HE GNN, a privacy-preserving graph inference framework that employs parallel Graph Isomorphism Network (GIN) and Graph Attention Network (GAT) branches followed by learnable gated fusion, with embedding-level homomorphic encryption for secure downstream classification. Graph embeddings are generated in plaintext and subsequently classified under Paillier and CKKS encryption schemes using a Logistic Regression classifier, enabling secure inference while reducing computational overhead. Experimental evaluation on the MUTAG, Cora, and PROTEINS benchmark datasets demonstrates that the proposed framework preserves predictive performance under encrypted inference with only negligible deviation from plaintext execution. The Hybrid-HE GNN achieved encrypted accuracies of 87.74%, 91.10%, and 87.12% on MUTAG, Cora, and PROTEINS, respectively, outperforming baseline GNN models across the evaluated datasets. Furthermore, CKKS batched inference substantially reduced latency compared with Paillier encryption while maintaining equivalent classification performance. These results demonstrate that embedding-level homomorphic encryption provides an effective balance between privacy preservation, predictive accuracy, and computational efficiency for graph neural network inference.