Abstract

Credit card fraud detection continues to be a major challenge in the financial industry due to extreme class imbalance, where fraudulent transactions occur far less frequently than legitimate ones. Traditional machine learning models often perform poorly on such imbalanced datasets, resulting in inadequate fraud detection rates. This paper introduces a sophisticated fraud detection framework that utilizes Variational Autoencoders (VAEs) to generate synthetic data and the Synthetic Minority Over-sampling Technique (SMOTE) to balance the minority class. Our hybrid approach generates realistic synthetic fraudulent samples while mitigating overfitting and loss of information associated with traditional oversampling techniques. We evaluated multiple classification models, including XGBoost, Deep Neural Networks (DNN), and CatBoost, using an augmented dataset and conducted a comparative analysis with conventional oversampling techniques. Extensive experiments demonstrate that our hybrid augmentation strategy significantly enhances fraud detection performance by increasing recall and F1-score while reducing false positives. We also discuss the trade-offs between different synthetic data generation techniques and their impact on classifier performance. Furthermore, we explore adversarial training techniques and their potential for real-time fraud detection deployment.