Credit Card Fraud Detection via Deep Learning Method Using Data Balance Tools

Previous studies have shown the effectiveness of deep learning algorithms in improving the detection of credit card fraud, which has become a major issue for financial institutions. This paper discusses how two deep learning methods, Convolutional Neural Network (CNN) and auto-encoder, perform in the fraud detection task using two different datasets. This research utilizes a brand-new dataset with all raw input variables and a Université Libre de Bruxelles (ULB) transaction dataset, which has been preprocessed with PCA technology. Since imbalanced datasets can affect the training quality, we further preprocess the datasets using random under-sampling and the Synthetic Minority Oversampling Techniques (SMOTE) to balance the datasets. Through the experimental results, we find that the networks perform well for the ULB dataset with 93% accuracy in prediction, but perform poorly for the independent input dataset, and the performance can be improved when increasing the complexity of the network. We also notice that the under-sampling method helps improve prediction accuracy better than the oversampling method. The results indicate that more complicated networks are required to detect fraud when the criterion for fraud is stricter, while balancing the dataset before training will improve the results.