Synthetic data augmentation for tabular data via deep learning methods

Payments and online transactions have become an integral part of our daily lives. Consequently, it is of paramount importance to defend against attacks from malicious users who engage in various types of fraudulent activities. Currently, most fraud detection approaches require a training dataset containing records of both benign and malicious usage. However, in practice, there are very few records of the latter activities, making it increasingly difficult to detect and prevent such rare frauds. This thesis focuses on the analysis of different augmentation techniques applied to three highly unbalanced datasets. The goal is to evaluate the performance of each model on tabular datasets and compare them with state-of-the-art machine learning techniques such as SMOTE, GMM, and oversampling. The evaluated models include GAN, WGAN-GP, CTGAN, TVAE and RTF. Each model's performance will be assessed based on Resemblance, Privacy, and Utility evaluations