Cross-agency financial fraud detection based on federated graph neural networks and privacy protection
TL;DR
This method constructs individual financial graphs using accounts, invoices, devices, terminals, and transaction relationships, and employs federated secure aggregation for joint modeling across organizations to address the discretization, isolation, and privacy protection issues in cross-organizational financial fraud detection.
Abstract
To address the discretization, isolation, and privacy protection issues in cross-organizational financial fraud detection, this paper proposes a heterogeneous graph neural network technique combining federated learning and privacy protection. This method constructs individual financial graphs using accounts, invoices, devices, terminals, and transaction relationships. Based on this, it utilizes relationship-aware message passing and temporal distance decay propagation to establish cross-organizational connections, and employs federated secure aggregation for joint modeling across organizations. Furthermore, differential privacy techniques, lightweight message transmission, and robust aggregation mechanisms are used to reduce the risk of information leakage and interference from anomalous updates. Experimental results demonstrate that this method achieves better accuracy, F1 score, AUC, and MCC, and maintains good robustness even under label noise, structural perturbations, and malicious edge injection. In addition, visual examples demonstrate that this model can provide interpretations of high-risk relationship paths and suspicious subgraphs.
