Machine Learning for Fraud Detection in Digital Banking and DeFi: Towards Sustainable Financial Systems

Abstract

The rapid development of digital banking and decentralized finance (DeFi) has increased the complexity of financial fraud detection. Although machine learning methods are increasingly applied to fraud detection, the literature remains fragmented regarding the comparative performance of different algorithms under highly imbalanced conditions and the implications of their application for decentralized and sustainable financial systems.The aim of this study is to comparatively evaluate selected supervised machine learning models for financial fraud detection and to examine the implications of their application for digital banking, DeFi, and sustainable financial systems.The study uses the publicly available Credit Card Fraud Detection dataset containing anonymized financial transactions with a highly imbalanced distribution between legitimate and fraudulent observations. Three supervised machine learning algorithms were comparatively evaluated: Logistic Regression, Random Forest, and XGBoost. Data preparation and model evaluation were implemented in Python using the Scikit-learn and XGBoost libraries. Model performance was assessed using precision, recall, F1-score, and ROC-AUC.The comparative results indicate that the ensemble-based models, particularly XGBoost, achieved stronger overall performance than Logistic Regression across the selected evaluation metrics. Random Forest also demonstrated competitive performance, supporting the applicability of ensemble approaches to fraud detection in complex and highly imbalanced financial data.The scientific contribution of the study consists in the comparative assessment of three widely used machine learning approaches within a unified fraud detection framework and in linking the obtained findings to the emerging challenges of DeFi and sustainable financial systems. The study also discusses the role of explainable artificial intelligence (XAI) in improving transparency and interpretability in financial machine learning applications.The findings may support financial institutions and digital financial service providers in selecting appropriate machine learning approaches for fraud detection and in balancing predictive performance with transparency and operational requirements. The discussion of DeFi highlights the need for adaptive analytical approaches capable of addressing the specific risks of decentralized financial ecosystems.The study concludes that machine learning, particularly ensemble-based approaches, has significant potential to strengthen fraud detection and contribute to safer, more resilient, and sustainable digital financial systems.