Authors: Chapekar Ved Sudhir, Prof. M. A. Auti, Prof. M. C. Kolse, Prof. S. K. Said
Abstract: Fraud in banking, digital payments, online shopping and insurance causes large financial losses and erodes customer trust. Traditional fraud detection relies on predefined rules and manual verification, which struggle with complex and constantly changing fraud patterns. Artificial Intelligence (AI) enables systems that analyse large volumes of transaction data in real time, recognize unusual behaviour and generate alerts. This paper reviews AI-based fraud detection. It explains the characteristics of fraud data, including extreme class imbalance, concept drift and delayed labels, and surveys statistical approaches, supervised classifiers such as logistic regression, random forests and gradient boosting, unsupervised anomaly detection, sequence models that learn from customers' transaction histories, and graph neural networks that expose fraud rings. Hybrid systems that combine supervised and unsupervised learning, benchmark datasets and evaluation metrics suited to imbalanced problems are discussed. A five-layer real-time fraud detection architecture is proposed, along with research gaps, objectives, a comparison of methods and challenges including false positives, explainability and adversarial adaptation by fraudsters. The review concludes that layered AI systems with human investigators in the loop provide the most effective defence.
DOI: https://doi.org/10.5281/zenodo.24219829
