Anti-confrontational Domain Data Generation Based on Improved WGAN

Abstract

The Domain Generate Algorithm (DGA) is used by a large number of botnets to evade detection. At present, the mainstream machine learning detection technology not only lacks the training data with evolutionary value, but also has the security problem that the model input sample is attacked. The Generative Adversarial Network (GAN) suggested by Goodfellow offers the possibility of solving the above problems, and WGAN is a variant of the GAN model implementation [1]. In this paper, an improved method for generating adversarial domain names by improved WGAN character domain name generator is proposed to improve model detection capability and expand effective training set. Experimental results show that this method produces adversarial domain names that are more consistent with human naming than traditional GAN models, adding these training sets with adversarial factors improves the discriminant hit ratio of the model to unknown domain names.