A hybrid deep learning framework using synthetic oversampling, autoencoder, convolutional neural networks, and an attention mechanism for credit card fraud detection

Credit card fraud is still a big problem for banks and other financial organizations throughout the globe. It hurts consumer confidence and financial stability. Despite significant progress in fraud detection, existing algorithms struggle with highly imbalanced datasets dominated by legitimate transactions. This article addresses this issue by proposing by suggesting a new way to solve it that combines the Synthetic Minority Oversampling Method (SMOTE), autoencoder, Convolutional Neural Networks (CNNs), and attention mechanism into one framework (SMOTE-AE-CNN-Att). The technique starts by utilizing SMOTE to balance the dataset, then uses AE-CNN-Att models to extract features, and then uses classic Machine Learning (ML) methods to classify the data. The suggested method has been shown to be very accurate (> 99.9%) in finding faket transactions while keeping important performance metrics like precision (up to 90.07%), recall (up to 91.13%), and F1-score (up to 90.60%). When compared to other techniques, the SMOTE-AE-CNN-Att model does a better job of finding a good balance between accuracy and recall, which is very important for finding fraud. This research shows how Deep Learning (DL) methods might make it much easier to detect fraud in credit card transactions. This would lead to better security and consumer protection in financial transactions.