Abstract
Domain generation algorithms (DGAs) are frequently employed by malware to generate domains used for connecting to command-and-control (C2) servers. Recent work in DGA detection leveraged deep learning architectures like convolutional neural networks (CNNs) and character-level long short-term memory networks (LSTMs) to classify domains. However, these classifiers perform poorly with wordlist-based DGA families, which generate domains by pseudorandomly concatenating dictionary words. We propose a novel approach that combines context-sensitive word embeddings with a simple fully-connected classifier to perform classification of domains based on word-level information. The word embeddings were pre-trained on a large unrelated corpus and left frozen during the training on domain data. The resulting small number of trainable parameters enabled extremely short training durations, while the transfer of language knowledge stored in the representations allowed for high-performing models with small training datasets. We show that this architecture reliably outperformed existing techniques on wordlist-based DGA families with just 30 DGA training examples and achieved state-of-the-art performance with around 100 DGA training examples, all while requiring an order of magnitude less time to train compared to current techniques. Of special note is the technique's performance on the matsnu DGA: the classifier attained a 89.5% detection rate with a 1:1,000 false positive rate (FPR) after training on only 30 examples of the DGA domains, and a 91.2% detection rate with a 1:10,000 FPR after 90 examples. Considering that some of these DGAs have wordlists of several hundred words, our results demonstrate that this technique does not rely on the classifier learning the DGA wordlists. Instead, the classifier is able to learn the semantic signatures of the wordlist-based DGA families.
Abstract (translated)
URL
https://arxiv.org/abs/1811.08705
