Protecting Contactless Credit Card Payments from Fraud through Ambient Authentication and Machine Learning

Abstract

Credit card fraud (CCF) is a persistent issue in the financial sector with serious consequences. Data mining has proven to be extremely useful in detecting fraud in online transactions. However, detecting CCF through data mining is quite a difficult task because of two causes: constant changes in the profiles of normal and fraudulent behaviour, and the highly skewed nature of the data sets. The outcome of fraud detection in credit card transactions depends on the sampling approach, detection techniques, and variable selection. This work studies the performance of K-Nearest Neighbor, Naive Bayes, Logistic Regression and Random Forest algorithms on a highly skewed dataset. The dataset contains 2,84,807 transactions and has been collected from European cardholder transactions. A hybrid of under-sampling and oversampling techniques has been used on the skewed data. The four techniques were utilized on both data namely preprocessed and raw, and the results are evaluated using specificity, accuracy, sensitivity, and F1-score. The outcomes show that the optimal accuracy for Naive Bayes, Logistic Regression, K-Nearest Neighbor and Random Forest classifiers are 98.72%, 52.34%, 96.89%, 91.67%, respectively. The comparative results indicate that K-Nearest Neighbor performs better than Logistic Regression, Random Forest and Naive Bayes techniques.