A comparative study of autoencoders and GANs for credit card fraud detection

Fraudulent behavior in the financial sector and its implications constitute a significant threat to the economy, society, and public trust. Detecting such behavior is therefore of crucial importance, as it enables timely intervention, and can potentially act as a deterrent. Despite this, fraud detection poses significant challenges, as it typically relies on limited labeled data and must operate under severe class imbalance. This thesis, using credit card transaction data, studies the fraud detection performance of autoencoders and GANs, while examining cross-model, architecture-level, and supervision-regime comparisons. By implementing a one-class, semi-supervised learning framework, where models are trained exclusively on non-fraudulent transactions, these deep learning models are employed to identify fraudulent transactions through various anomaly scoring and thresholding strategies, while exploring different optimization frameworks. Model performance is evaluated using the F1-score and confusion matrices, which are particularly appropriate for rare-event classification problems. The empirical analysis carried out in this study indicates that autoencoders exhibit stronger classification performance than GANs. It further shows that TPE-based optimization can yield better classification results when combined with the use of fraud labels during the optimization stage, compared to semi-supervised TPE optimization and manual search, highlighting the importance of hyperparameter selection. Overall, the findings emphasize the practical relevance of deep learning models, one-class training paradigms, and optimization method choice for fraud detection in highly imbalanced financial datasets.