Improving Credit Card Fraud Detection Through Transformer-Enhanced Gan Oversampling

The issue of credit card fraud detection is also not easy because the ratio of legitimate and fraudulent activities is very high, and thus limits the ability of the standard classifier to learn. The traditional oversampling methods, like SMOTE, produce linearly interpolated samples of the minority that do not represent the more intricate and nonlinear associations of fraudulent behavior. Even more recent tabular generative models, including CTGAN and TVAE, are better at improving sample diversity and still fail to capture the interactions between high-dimensional features and may collapse modes in even sparse minority classes.,To overcome these shortcomings, this paper will introduce a new hybrid Transformer-Enhanced Generative Adversarial Network (T-GAN) targeted at tabular fraud data. In contrast to the current CTGAN- or VAE-based architectures, the proposed one incorporates a Transformer encoder in the generator along with the feature recalibration of the Squeeze-and-Excitation (SE) feature improvements, which allows the model to capture long-range feature interactions without losing interference of attributes. The T-GAN design can generate large-scale and structurally coherent synthetic fraud samples and effectively address the imbalance problem.,On actual data of Credit Card Fraud Detection, experiments show that T-GAN notably beats SMOTE, CTGAN and TVAE on several classifiers- including Logistic Regression, Random Forest, XGBoost and SVM- with the domain improvements in Recall and F1-score as well as AUC. These findings support the fact that Transformer-guided adversarial synthesis is much more effective in promoting minority representation and offering a more trustworthy baseline to the downstream fraud detection models.