基于生成式对抗网络的恶意URL数据生成与检测

Generating and Detection Malicious URL Based on Generative Adversarial Networks

摘要: 针对基于机器学习的恶意网页识别中对数据集的收集和标注敏感的问题，提出了一种基于生成式对抗网络(GAN)的检测方法，并且设计了编码器，将恶意URL进行字符级编码。通过使用少量样本训练模型，通过GAN拟合真实样本的能力，生成恶意网页样本。本文在传统GAN的基础上增加了一个判别器用来判别良性和恶性网页，达到了判别恶意网页的作用。最后通过横纵对比实验，分别验证了生成数据的可行以及判别模型可以达到当前有监督分类器相当的效果。

Abstract: Malicious web page recognition based on machine learning is sensitive to data collection and annotation. This paper proposes a method of generating and detecting malicious web pages based on Generative Adversarial Networks (GAN). Design an encoder in order to encode malicious URL at character level. A small number of samples were used to train the model, and the ability of GAN to fit real samples was used to generate malicious web page samples. On the basis of traditional GAN, this paper adds a discriminator to discriminate benign and malignant web pages, and achieves the function of discriminating malicious web pages. Finally, the feasibility of the generated data and the effectiveness of the discriminant model with the currently supervised classifier are verified by vertical and horizontal comparison experiments.