A Novel DGA Domain Adversarial Sample Generation Method By Geometric Perturbation

Malicious domain names detection is an important technology in network security. Attackers mainly use domain generation algorithms (DGAs) to carry out malicious network attacks. Although DGA domain detection based on deep learning has good performance, recent studies have shown deep learning methods are vulnerable to adversarial examples. Therefore, we focus on the generation of DGA domain adversarial samples. In this paper, firstly we introduce the concept of geometric vectors into the adversarial samples and prove the effectiveness of the attack from the perspective of mathematical geometry. Secondly, we propose an algorithm of DGA domain adversarial sample generation based on the geometric perturbation, which uses the method of geometric vector to generate adversarial perturbation and adds it to DGA malicious domain name data to generate adversarial samples. To further verify the effectiveness of our algorithm, four DGA domain detection classifiers are used to test the generated adversarial samples, and the experimental results show that the classifiers are not able to resist the attacks of our method. Compared with other DGA domain adversarial sample generation methods, the proposed method has better performance.