Cross-Institution Cooperative Detection Model of Insurance Fraud Based on Graph Neural Networks and Federated Learning
Year: 2026
Type: article
Abstract: The core challenge in cross-institution insurance fraud detection is that fraud-ring connectivity crosses institutional boundaries, while privacy regulations prohibit the centralized aggregation of raw claims data. This paper proposes CFGFraud, a federated graph detection framework for partially observable graphs. In this framework, a virtual-node placeholder mechanism makes the local model aware of cross-institutional hub entities; institutional attention mask separates the parameterization of real and virtual connections and imposes a learnable attenuation coefficient; and a federated attention alignment strategy constrains drift in each institution’s attention pattern during aggregation. The non-IID experiment with five clients on the IEEE-CIS dataset shows that CFGFraud achieves an AUPRC of 0.791, 0.057 higher than FedAvg+GAT. Fraud-ring recall reaches 0.723, 24.4% higher than FedAvg+GAT. The standard deviation across clients is compressed from 0.029 to 0.012. Single-round computation costs increase by about 8%, while communication overhead increases by less than 0.01%. Ablation experiments show that the virtual-node mechanism accounts for 31.1% of the fraud-ring recall gain, and collaboration gain is positively correlated with entity overlap across institutions. This framework provides a deployable technical path for cross-agency relational fraud detection under the constraint of data confinement.
Source: Journal of Digital Frontier
DOI: 10.67541/jdf2613
Institution: Harvard University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235379414
