Domain Generation Algorithms (DGAs) are a popular technique used by contemporary malware for commandand-control (C&C) purposes. Such malware utilizes DGAs to create a set of domain names that, when resolved, provide information necessary to establish a link to a C&C server. Automated discovery of such domain names in real-time DNS traffic is critical for network security as it allows to detect infection, and, in some cases, take countermeasures to disrupt the communication and identify infected machines. Detection of the specific DGA malware family provides the administrator valuable information about the kind of infection and steps that need to be taken. In this paper we compare and evaluate machine learning methods that classify domain names as benign or DGA, and label the latter according to their malware family. Unlike previous work, we select data for test and training sets according to observation time and known seeds. This allows us to assess the robustness of the trained classifiers for detecting domains generated by the same families at a different time or when seeds change. Our study includes tree ensemble models based on human-engineered features and deep neural networks that learn features automatically from domain names. We find that all state-of-the-art classifiers are significantly better at catching domain names from malware families with a time-dependent seed compared to time-invariant DGAs. In addition, when applying the trained classifiers on a day of real traffic, we find that many domain names unjustifiably are flagged as malicious, thereby revealing the shortcomings of relying on a standard whitelist for training a production grade DGA detection system.

No abstract

Domain generation algorithms (DGAs) have become commonplace in malware that seeks to establish command and control communication between an infected machine and the botmaster. DGAs dynamically and consistently generate large volumes of malicious domain names, only a few of which are registered by the botmaster, within a short time window around their generation time, and subsequently resolved when the malware on the infected machine tries to access them. Deep neural networks that can classify domain names as benign or malicious are of great interest in the real-time defense against DGAs. In contrast with traditional machine learning models, deep networks do not rely on human engineered features. Instead, they can learn features automatically from data, provided that they are supplied with sufficiently large amounts of suitable training data. Obtaining cleanly labeled ground truth data is difficult and time consuming. Heuristically labeled data could potentially provide a source of training data for weakly supervised training of DGA detectors. We propose a set of heuristics for automatically labeling domain names monitored in real traffic, and then train and evaluate classifiers with the proposed heuristically labeled dataset. We show through experiments on a dataset with 50 million domain names that such heuristically labeled data is very useful in practice to improve the predictive accuracy of deep learning-based DGA classifiers, and that these deep neural networks significantly outperform a random forest classifier with human engineered features. INDEX TERMS Deep learning, random forest, text classification, heuristically labeled data, domain generation algorithms, cybersecurity, command and control.