Federated Financial Fraud Detection and Investigation Platform using Machine Learning

The rapid globalization of electronic banking, real-time payment switches, and mobile point-of-sale gateways has drastically amplified the attack surface for financial fraud. Intercepting fraudulent transactions in production environments presents severe operational hurdles driven by extreme data skewness—where illicit transactions constitute less than 0.2% of total traffic—and stringent legal firewalls, including the General Data Protection Regulation (GDPR) and the Payment Card Industry Data Security Standard (PCI-DSS), which prohibit commercial institutions from sharing proprietary financial ledgers into centralized repositories. Furthermore, conventional black-box classification models fail to provide explainable decision paths, creating high false-positive alert fatigue and crippling operational compliance workflows. To address these systemic challenges, this research proposes an enterprise-grade, privacy-preserving Federated Financial Fraud Detection and Investigation Platform. The architecture establishes a decentralized federated learning ring orchestrated via the Flower framework utilizing Federated Averaging (FedAvg) across simulated institutional nodes (Bank A, Bank B, and Bank C). In parallel, live transactions are ingested through an optimized 776-dimensional sparse preprocessing pipeline that captures nonlinear behavioral traits, cyclic temporal dynamics, and customer-to-merchant Haversine geodetic distances. In-flight scoring is executed through a heterogeneous multistage soft-voting ensemble combining Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), CatBoost, and Random Forest. Model decisions are coupled with an Explainable AI (XAI) attribution engine powered by Tree SHAP, yielding localized, additive feature attributions for root-cause auditability. Evaluated on a comprehensive empirical benchmark of 555,719 real-world credit card transactions, the soft-voting ensemble achieves a 94.83% fraud recall, 98.87% overall classification accuracy, and an Area Under the Precision-Recall Curve (PR-AUC) of 0.7500, successfully capturing 2,034 out of 2,145 fraudulent events while maintaining an ultra-low inference latency of 0.0749 ms per transaction. The privacy-preserving federated baseline achieves 71.00% fraud recall and 0.8891 ROC-AUC with zero cross-institutional data disclosure. The platform is realized through a production-ready stack integrating Python FastAPI, Java Spring Boot 3.5, an immutable PostgreSQL audit ledger, and a responsive React triage console.