Federated Transaction Graph Analysis for Regulatory Compliant Banking Fraud Detection
Main Article Content
Abstract
Fraudulent transactions often involve networks of accounts, customers, devices, merchants, beneficiaries, and financial institutions. An individual bank may therefore observe only a limited portion of a coordinated fraud scheme. Although centralizing transaction data from multiple banks could improve fraud detection, such centralization may conflict with privacy requirements, confidentiality obligations, cybersecurity controls, data-localization rules, and restrictions on crossborder processing. This paper proposes a conceptual federated graph intelligence architecture for collaborative cross-bank fraud detection. Each participating bank constructs and retains a local heterogeneous temporal graph containing customers, accounts, transactions, devices, merchants, and beneficiaries. Compatible graph neural networks are trained locally, while protected model updates are communicated through secure aggregation. Where permitted by law, an independent privacy-preserving entity-matching service provides narrowly defined pseudonymous overlap signals without transferring raw identifiers or complete transaction graphs. The framework separates payment-fraud detection, muleaccount identification, fraud-ring discovery, and anti-money-laundering analysis as distinct but related tasks. It also distinguishes bank-level differential privacy from transaction-, edge-, node-, customer-, and subject-level protection. Secure aggregation, contribution clipping, differential privacy, temporal controls, robust validation, explanation mechanisms, human review, and governance controls are treated as complementary safeguards. Because the paper proposes a conceptual architecture rather than an implemented system, it does not report experimental performance. Instead, it presents an evaluation framework covering predictive utility, privacy leakage, calibration, fairness, adversarial robustness, latency, communication cost, and auditability. The analysis concludes that federated graph intelligence may improve cross-bank fraud detection, but it cannot independently establish regulatory compliance. Deployment requires an explicit legal basis, clearly defined institutional responsibilities, privacy-impact assessment, common data standards, independent model validation, security testing, incident management, and meaningful human oversight.