Federated graph neural network for privacy-preserved supply chain data sharing

Abstract

Machine learning plays an increasingly important role in supply chain management. Due to privacy and security concerns, enterprises are reluctant to share their raw data, which leads to missing links in supply chains. To address privacy issue and promote data sharing in supply chain, we propose a new federated graph neural network named Isomorphic Federated Graph Neural Network (IFGNN) for supply chain data sharing. IFGNN consists of a server and multiple clients. The server is a lightweight parameter server with an efficient parameter updating algorithm. A client is assigned to each node in the supply chain network. Every supplier client is linked to its first-order neighbors, which means they have supply–demand relationship. The topology of the input network is identical to that of the supplier clients. Experimental results on a newly collected vehicle supply chain dataset show that the performance of IFGNN is close to its centralized counterpart. This work demonstrates that it is feasible to protect supply chain data privacy without a significant loss of prediction accuracy. Federated learning provides a new solution for promoting data sharing and collaborative machine learning in supply chain.