Real-Time Anomaly Detection Using ML, DL, and XAI for Financial Fraud: A Systematic Review

Abstract

Digital payments, fintech, and online banking are now part of our daily lives. Making a transfer, paying for a service, or shopping online is faster, but this progress has also given rise to new forms of fraud. Rule-based systems remain useful when the pattern is already known; the problem arises when attackers change their tactics. For this reason, this study reviews the use of machine learning, deep learning, and Explainable Artificial Intelligence to detect anomalies and prevent fraud in digital payment systems. The research was conducted as a systematic literature review. PICOC helped define the scope and search terms, while PRISMA guided the selection of articles. Scopus, Web of Science, and ScienceDirect were consulted. After applying the inclusion, exclusion, and methodological evaluation criteria, 33 studies were retained. The articles addressed cases of card fraud, digital banking, blockchain, money laundering, and other suspicious transactions. The most frequently mentioned models were Random Forest, XGBoost, SVM, Isolation Forest, LSTM, CNN, autoencoders, and graph networks. Hybrid approaches and tools such as SHAP and LIME also appeared, used to show which variables influenced an alert. Several studies reported good results in terms of accuracy, recall, and F1 score. However, challenges remain regarding synthetic data, class imbalance, limited information on latency, and insufficient validation with financial specialists. Their performance still needs to be tested with real-world data and under conditions similar to those of a financial institution.