Abstract: This study looks into two main areas related to detecting fraud without using labeled data. The first area focuses on using evaluation methods from Generative Adversarial Networks (GANs) to spot fraud or outliers by calculating the differences between samples from normal (non-fraudulent) data and generated data. The second area explores whether the discriminator of a GAN can be used as a feature space for calculating these differences. Since fraudulent examples are not available during training, the problem is approached as finding outliers.