Leveraging n-gram neural embeddings to improve deep learning DGA detection

Several families of malware are based on the need to establish a connection with a Command and Control (C&C) server. In addition, to avoid detection, these servers "hide" behind domain names that are periodically changed according to a specific Domain Generation Algorithm (DGA). Hence, the malware that has infected a particular host uses the same DGA to make DNS queries in order to establish a connection with the C&C server. The identification of "malicious" domain names used in DNS queries is therefore crucial for their detection. For this purpose, various machine learning techniques have been used, in particular, recently, deep learning techniques have proved especially effective. However, to get good results, these techniques require very large and labelled training datasets. Nevertheless, the construction of such datasets, decidedly with regard to the collection of malicious domain names, is a very difficult and nonscalable task. In this paper, therefore, we explore the possibility of exploiting unsupervised character n-gram embeddings to improve the performance of a Deep Learning DGA classifier. Embeddings are trained using a large dataset of benign names, opening up the possibility of using a small classifier training dataset requiring a small number of malicious names. A series of experiments, which use the same embedding for classifiers trained with datasets of increasing size, are then presented. These experiments show how the embedding is particularly effective for classifiers trained with small datasets having a small number of malicious names.