Enhanced Credit Card Fraud Detection using Data Augmentation and Deep Anomaly Detection Techniques

The increasing prevalence of credit card fraud presents significant risks for financial institutions and consumers due to extreme class imbalance in transaction data. This thesis investigates the impact of data augmentation on fraud detection, comparing Synthetic Minority Over-sampling Technique (SMOTE) and Conditional Tabular Generative Adversarial Networks (CTGAN). A novel "balanced augmentation"strategy, equally upsampling fraud and non-fraud classes, was assessed using Random Forest, Artificial Neural Network (ANN), Compensated ANN (C-ANN), and Support Vector Machine (SVM). Rigorous evaluations involving cross-validation and threshold tuning highlighted that balanced augmentation substantially improved fraud detection performance, particularly enhancing recall from baseline levels (~75%) to near-perfect (~99.98%) with CTGAN. Random Forest combined with SMOTE delivered optimal overall reliability and stability, achieving the highest practical performance (AUC = 99.99%) under realistic conditions. Findings underscore the critical role of balanced data augmentation in operational fraud detection systems, providing clear, practical guidance for financial institutions.