Temporal transaction network anomaly detection for Industrial Internet of Things with federated graph neural networks

Abstract

The Industrial Internet of Things (IIoT) has experienced significant advancements in recent years, resulting in a considerable increase in the volume of data generated by interconnected devices. This surge in data has created new opportunities to enhance the quality of service in machine learning applications within the IIoT through data sharing. Among these applications, anomaly detection in transaction networks utilizing graph neural networks (GNNs) has emerged as a prominent research topic. However, most current anomaly detection methods either focus exclusively on single-faceted transaction information or assume that multiple types of transaction network data are centrally stored or shared. In the field of IIoT scenario, privacy concerns and legal restrictions frequently hinder data centralization, resulting in data islands, which refer to decentralized multisource transaction information. Therefore, we propose a novel federated GNNs framework for the temporal transaction network anomaly detection, designated as FedGT${}^2$AD. Specifically, the training process is bifurcated: client-side privacy temporal transaction network feature extraction is conducted locally at its corresponding client, while privacy-protected feature aggregation from all clients occurs on a trusted server. To facilitate more effective anomaly detection, each client initially models edge features and temporal transaction information as node attributes, along with network snapshots for subsequent graph feature computation with GNNs. During the integration process, the server integrates the node-level embedding and computes multisource transaction network features from all clients following a differential privacy mechanism to ensure client-side data security. The experimental results on three decentralized multisource transaction networks demonstrated that FedGT${}^2$AD outperforms baseline methods by 0.9% to 2.7% in accuracy. Overall, FedGT${}^2$AD offers a promising approach for mining decentralized multisource transaction networks while preserving privacy in anomaly detection tasks.