A Fraudulent Data Simulation Method Based on Generative Adversarial Networks

Abstract

In view of the positive and negative sample imbalance in the process of establishing anti-fraud rules and models of the third-party payment, this paper simulated the fraudulent transactions by means of generative adversarial networks. In the design of the model, the output of the discriminator is optimized where classification results are set to three categories; the loss function of the GAN model is changed to include two parts: the source loss function and the category loss function. The generated data and the real business data are mixed according to a certain ratio, to train a fraud detection model. Verify the generation effect of this simulation method by comparing the detection effects of different detection models.