Conjugacy Search Problem Homomorphic Encryption Based Vertically Federated Graph Neural Network
Bo Mi, Ran Zeng, Ling Zhao, Siyuan Zeng, Fuyuan Wang, Qi Zhou · 2024
The advent of graph neural networks has enabled neural networks to utilize graph-structured data and attributes. In real life, there is a large amount of graph-structured data, such as social networks, traffic networks, etc. However, the superior performance of graph neural networks requires extensive information about nodes and edges, which may be distributed across different clients. In this paper, we propose a Vertical Federated Graph Neural Network (VFGNN) that combines federated learning and split learning to protect users’ local data and break information silos while improving the accuracy of the model. Specifically, we divide the computation graph into two parts, each client performs calculations on a front-end fixed part of the computational graph using local data and forwards its results to the server, which performs aggregation and further operations. For the sake of efficiency and privacy, we use an oblivious pseudo-random function based on cuckoo hashing to perform privacy set intersection during data preprocessing. A homomorphic encryption scheme based on the Conjugacy Search Problem (CSP) is used to encrypt information that needs to be sent to the server to prevent potential information leakage.