"Artificial Intelligence in Banking Fraud Detection: A Systematic Review using the Prisma Approach"

Abstract

The rapid growth of digital banking, instant payments, mobile applications and card-not-present transactions has improved access to financial services while simultaneously widening the attack surface for fraud.Conventional rulebased systems remain indispensable for regulatory control, yet they are often unable to recognize subtle, coordinated and previously unseen patterns.This paper presents a condensed systematic review of artificial intelligence (AI) in banking fraud detection, organised according to the PRISMA 2020 reporting logic.It synthesizes evidence on supervised learning, anomaly detection, deep learning, graph analytics, natural language processing and hybrid decision systems.The review shows that tree-based ensembles generally perform strongly on structured transaction data; sequential neural networks are useful for behavioural histories; graph methods are promising for mule-account and fraud-ring detection; and anomaly models support discovery where labels are scarce.However, accuracy alone is misleading in highly imbalanced datasets.Precision, recall, F1-score, precision-recall area, false-positive cost, detection latency and financial value saved are more meaningful.The review further identifies concept drift, delayed labels, data leakage, limited public datasets, explainability, privacy and bias as major implementation concerns.It concludes that the most credible banking architecture is layered, adaptive and human-supervised rather than dependent on a single algorithm.PRISMA 2020 is used as the reporting framework because it improves transparency in identification, screening, eligibility assessment and synthesis of studies.