Blockchain-Enhanced Secure Data Sharing in Financial Institutions: A Federated Learning Framework for Privacy-Preserving Analytics
TL;DR
The proposed framework for financial system fraud detection that is safe and protects privacy while resolving issues with data sharing, legal restrictions, and cybersecurity threats is appropriate for practical financial applications since it successfully improves fraud detection while guaranteeing Privacy Preservation, security, and openness.
Abstract
Purpose The goal of this research is to create a framework for financial system fraud detection that is safe and protects privacy while resolving issues with data sharing, legal restrictions, and cybersecurity threats. To facilitate cooperative model training without exchanging private raw data. Methods A blockchain-enhanced Federated Learning (FL) method is suggested. The technique divides a centralized dataset among several clients to simulate an FL environment. Local training is done using a TabNet deep learning model, and hyperparameter tweaking is done via Bayesian optimization. Input features are normalized and passed through TabNet's sequential feature transformer and attentive transformer, which generate sparse attention masks to select the most discriminative attributes at each decision step; the aggregated decision representations are then classified using a softmax-based prediction layer. Model updates are transmitted using AES-256 encryption to guarantee data security, and their integrity is confirmed by SHA-256 hashing. Blockchain technology provides transparency and tamper resistance by recording model update hashes via Hyperledger Fabric. The integration of AES-256 encryption, SHA-256 integrity verification, and Hyperledger Fabric blockchain enables secure peer-to-peer model transmission, making the proposed framework well suited for decentralized financial environments. Results With an accuracy of 0.9924, precision of 0.9947, recall of 0.9910, and low rates of false positives and false negatives, the findings show excellent performance. Validation accuracy improved consistently across successive federated communication rounds, rising from 0.9887 to 0.9933, confirming stable model convergence, while the framework achieved a ROC-AUC of 0.9980 and PR-AUC of 0.9986. The suggested architecture performs better than conventional models like decision trees and logistic regression. Conclusion The framework is appropriate for practical financial applications since it successfully improves fraud detection while guaranteeing Privacy Preservation, security, and openness.