Privacy-Preserving Federated Learning for Cross-Institutional Fraud Detection and Financial Risk Intelligence

Abstract

Financial fraud increasingly spans institutions rather than staying inside one: a synthetic identity is used to open accounts at several lenders at once, a money-mule network moves funds through accounts at multiple banks, and a loan-stacking scheme draws on the fact that no single lender sees an applicant’s simultaneous exposure elsewhere. Any one institution’s fraud model, trained only on its own data, is structurally blind to the cross-institutional part of the pattern, yet the obvious remedy pooling raw transaction and applicant data across institutions into one shared dataset is foreclosed by bank secrecy obligations, competitive sensitivity, and data-protection law, all of which the peer-reviewed information-sharing literature treats as substantial and legitimate constraints rather than mere friction. Federated learning has been proposed as a way through this impasse: institutions train a shared model collaboratively without ever exporting their raw data, and a growing peer-reviewed literature demonstrates the approach’s technical feasibility, including specific applications to credit card and financial fraud detection. That same literature, however, establishes that a naive federated deployment is not automatically safe or robust: model updates alone can leak sensitive information without additional privacy-preserving mechanisms, the technique is vulnerable to poisoning by a malicious participant, and statistical heterogeneity across institutions’ data complicates convergence. This paper addresses the resulting gap by developing the Federated Risk Intelligence Network (FRIN): a reference architecture in which participating institutions train locally, apply differential privacy to their updates before those updates ever leave the institution, submit them through a secure aggregation protocol that prevents any single party from inspecting another’s contribution, and receive back a shared fraud and risk model screened for adversarial manipulation. Because no production deployment of this specific consortium architecture exists to evaluate, the paper adopts a design-science approach: the framework is built from and evaluated against peer-reviewed findings in federated learning fundamentals, privacy-preserving machine learning, federated learning’s statistical and adversarial challenges, federated learning applications in financial fraud detection, and the economics of cross-institutional information sharing, and the quantitative comparison presented one comparative table and three  figures is explicitly labeled as a literature-grounded  estimate rather than a measured result. The paper’s contribution is a synthesis of five otherwise separate literatures into a single consortium-level architecture, and a candid account of what privacy-preserving federated learning can plausibly offer cross-institutional fraud detection, and what it cannot.