Adversarial DGA Domain Examples Generation and Detection

Abstract

Botnets have long relied on the Domain Generation Algorithm (DGA) to survive to this day. The detection rate of the DGA detection methods based on machine learning is already high. However, the models trained by the existing data sets sometimes are blind to new variant domains.To mitigate such problem, a method based on generation adversarial networks (GAN) called DnGAN is proposed to generate adversarial DGA examples in this paper. Experiment results show that the adversarial examples can effectively escape the detection of multiple detectors. And by using these adversarial examples as training data can effectively enhance the ability of the detector to identify DGA families that have not been seen before.