Hybrid Generative Adversarial Network Assisted Oversampling for Credit Card Fraud Detection

Abstract

Credit-card fraud detection is a challenging task due to the highly imbalanced nature of real-world transaction data, where fraudulent cases constitute a tiny fraction of the total. In this study, we address this challenge using the publicly available Credit Card Fraud Detection dataset from Kaggle, which contains 284,807 transactions, of which only about 0.17% are labeled as fraudulent. To ensure a realistic evaluation, the dataset is split chronologically, with the last 20% reserved as a hold-out test set, and the remaining 80% split into a training and validation set, with the final 10% used for validation. Feature preprocessing is performed using only the training set statistics to prevent information leakage. We propose a hybrid data-balancing framework that integrates a Generative Adversarial Network (GAN) to generate realistic synthetic fraud samples, followed by the Synthetic Minority Oversampling Technique (SMOTE) to achieve a target minority-to-majority ratio. This approach is designed to improve minority representation while limiting synthetic noise through GAN early stopping and balanced sample generation. Classifiers are trained on the resampled training set, with optimal decision thresholds determined from validation data to maximize the F1-score. Experimental results show that the best Baseline without oversampling —a Random Forest model with class weighting —achieves an accuracy of 0.999596, precision of 0.9333, recall of 0.7467, F1-score of 0.8296, ROC-AUC of 0.9556, and PR-AUC of 0.8037. Our proposed hybrid method, GAN+SMOTE with a target ratio of 0.35, combined with an unweighted Random Forest, improves recall to 0.7733 and ROC-AUC to 0.9762 while maintaining a comparable PR-AUC of 0.8003, though with slight reductions in precision and F1-score compared to the weighted Baseline. These results demonstrate that the hybrid oversampling strategy enhances ranking performance and minority-detection rates, particularly in scenarios where recall is critical. The complete implementation, including data preprocessing, GAN training, hybrid oversampling, model evaluation, and visualization scripts, is publicly available at https://github.com/yasserramadan202025/gan_smote to ensure full reproducibility and facilitate further research.