Online Fraud Transaction Detection Using Machine Learning

Abstract

The recent trends in online financial transactions is leading to the increase of fraudulent behavior, necessitating the development of more efficient and faster fraud detection techniques. The existing machine learning approaches experience challenges in dealing with the imbalance transaction dataset and fraud evolution. This paper describes the extended version of a one-dimensional Convolutional Neural Network (CNN1D) for detecting fraudulent transactions by learning spatial and temporal patterns from transaction data. The proposed approach consists of data preparation, feature extraction, imbalance management, benchmarking of the baseline models, CNN1D-based classification, and real-time fraud detection. Naïve Bayes and XGBoost algorithms are selected as benchmarking models for the proposed model evaluation. From the experimental results, the performance of the proposed approach is shown with the CNN1D accuracy of 98.16%, precision of 71.56%, recall of 62.22%, and F1-score of 65.53%, whereas the XGBoost algorithm provides the accuracy of 98.38% with high precision but lower recall. The response time of the CNN1D model is one second.