Adversarial Sample of Domain Generation Algorithm Based on Variational Autoencoder

This paper conducts an in-depth discussion on the attack and defense issues of DGA domain names in the adversarial sample environment. Based on the method of splitting benign domain names into <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$n$</tex>-gram segments (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$n$</tex> is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1,2,3,4$</tex>) and embedding semantic information, a high-quality n-gram dictionary is constructed, which provides a data basis for the analysis of the generation rules of large-scale benign domain names and various DGA domain names. On this basis, a framework for adversarial DGA domain name generation that comprehensively utilizes variational autoencoders (VAE), multi-head self-attention Transformer encoders and LSTM models is designed to generate TLVDGA with high concealment and strong adversarial properties. In order to alleviate the posterior collapse problem caused by KL divergence in VAE training, this study introduces the S-type KL annealing strategy. For the generated adversarial samples, this paper also conducts systematic attack experiments on two DGA domain name detection models with different architectures and multiple types of DGA domain names. Among them, TLVDGA achieved <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$69 \%, 63 {\%}$</tex> and 66% in KL, ED and JI distances respectively; in the attack experiment, the detection model's precision was reduced to 55%, the recall rate was reduced to 59%, and the false alarm rate was as high as 32%.