Mobile Payment Fraud Detection and Prevention System

Abstract

As mobile payment platforms continue to gain widespread adoption, the risks associated with fraudulent transactions have escalated, prompting the need for robust, real-time fraud detection systems. This thesis presents a machine learning-based framework to detect fraudulent activity in mobile payment transactions, with a focus on logistic regression (LR) and multi-layer perceptron (MLP) models. The study explores the efficacy of these models in identifying anomalies using a feature-rich dataset that includes temporal, transactional, and geolocation-based attributes. Preprocessing techniques such as normalization, encoding, and handling of class imbalance through SMOTE and ADASYN were employed to enhance model performance. Logistic regression offered strong explainability and low-latency predictions, making it suitable for regulated financial environments where auditability and interpretability are critical. On the other hand, the MLP model, trained using DeepLearning4J, demonstrated superior detection accuracy due to its ability to capture complex non-linear patterns, albeit with greater computational requirements. The research further investigates key features that differentiate fraudulent from legitimate transactions, including transaction amount, timestamp anomalies, and location mismatches. Techniques such as SHAP and LIME are discussed to improve the interpretability of the MLP model, and scalability strategies involving model compression, batch inference, and containerized deployments are outlined for real-time production readiness. Benchmarking experiments show that both models achieve acceptable latency for mobile payment environments, with the MLP model achieving higher F1 scores. The findings underscore the trade-offs between explainability, scalability, and predictive accuracy in the design of fraud detection systems. This study presents a practical methodology for deploying machine learning models within mobile payment ecosystems and provides a foundation for future enhancements, including the integration of streaming architectures and real-time user feedback mechanisms to support continuous learning and improvements.

Date of Award	2025
Original language	English
Supervisor	Erik Sonnleitner (Supervisor)