Adaptive Synthetic Data Generation and Integration for Enhanced Fraud Detection in Low-Resource Financial Domains

Synthetic data creation has become an increasingly important method for strengthening machine learning datasets, particularly in situations where privacy regulations, limited access, or insufficient diversity may hinder model effectiveness. This research examines the application of Conditional Tabular Generative Adversarial Networks (CTGAN) for producing synthetic credit card transaction data within a balanced binary classification setting. The original dataset consisted of an equal split between fraudulent and legitimate transactions, eliminating the requirement for class balancing. Using this dataset, CTGAN was trained to generate 1,000 synthetic records designed to replicate the statistical properties of the real data. To evaluate performance, three dataset variations were constructed: (i) the original data, (ii) a fully synthetic version, and (iii) a mixed dataset combining both real and generated samples. A Random Forest classifier was applied to each version, and results were measured using Accuracy, Precision, Recall, F1-score, and AUC-ROC. The findings revealed that synthetic data produced by CTGAN achieved comparable results to the real dataset, while the combined dataset produced small but consistent improvements in accuracy and recall. These results demonstrate that CTGAN can successfully model balanced financial datasets and create realistic synthetic records that maintain predictive accuracy.