Machine Learning-Based Fraud Detection in Digital Financial Transactions

Abstract

Digital payments, mobile banking, and instant transfers have increased transaction volumes and made financial services
easier to use. They have also given fraudsters new, faster, and harder-to-trace channels. Rule-based fraud controls cannot keep
up with the scale of digital transactions or with fraud patterns that change constantly. This review synthesizes research on
machine learning (ML) for fraud detection in digital financial transactions, organized around three themes: (1) the detection
techniques used, including supervised, unsupervised, hybrid, sequence-based, and graph-based models, and the methods
used to handle extreme class imbalance; (2) the operational requirements for running these models in production, including
real-time inference, zero-downtime model updates, concept drift management, explainability for investigators, and workflow
and compliance integration; and (3) the emerging use of large language models (LLMs) and autonomous agents in fraud
operations, together with the prompt engineering, guardrails, and audit mechanisms they require. The evidence shows that
ML models, particularly gradient-boosted ensembles, sequence models, and graph-based approaches, outperform static
rules on detection. Their effectiveness in practice, however, depends on handling delayed and noisy labels, adapting to drift,
keeping false positives manageable, and meeting regulatory demands for explainability and auditability. The review proposes
a layered framework for ML-based fraud management and identifies directions for future research.