Variants of Generative Adversarial Networks for Credit Card Fraud Detection

Abstract

One of the major problems in the field of Banking, Financial Services and Insurance (BFSI) is detecting fraudulent transactions. It is a big challenge to accurately detect fraudulent transactions because of the huge variation in the number of fraudulent and non-fraudulent samples, called the class imbalance problem. Many approaches address this problem, such as over-sampling, under-sampling, cost-sensitive methods, etc. to name a few. In recent years, Generative Adversarial Network (GAN)-based approaches for oversampling have drawn attention from both industry and academia to overcome this problem. In this paper, we have explored and compared various state-of-the-art GAN based approaches such as Vanilla GAN, Conditional GAN (cGAN), Wasserstein GAN (WGAN), and Conditional WGAN (WcGAN) for the problem of credit card fraud detection. It is found that the recent techniques have been very sensitive to the hyperparameters. To solve this problem we introduced two new methods, WGAN with Gradient Penalty (WGAN-GP) and Conditional WGAN with Gradient Penalty (WcGAN-GP) for credit card fraud detection. It is found that these approaches not only generate more realistic data but also provide more stable results.