Hardening Inline DGA Classifiers Against Adversarial Attacks
MetadataShow full item record
Domain Generation Algorithms (DGAs) are widely used by cybercriminals to generate domain names on-the-go for C&C (command-and-control) purposes of establishing communication with the bots and instructing them to perform malicious activities. It is therefore important to detect domains generated by DGAs to block the communication between the bot and C&C. In recent years, Machine Learning based DGA detection systems are widely used to address this problem. However, it is found that classifiers that rely only on the domain name to detect DGAs are highly vulnerable to adversarial attacks. Adversarial attacks are intentionally devised by an attacker to fool a classifier and cause it to produce erroneous results. This is a serious concern as it degrades the performance of DGA detection classifiers. In this thesis, we aim to defend DGA detection classifiers against adversarial attacks, without compromising the performance of existing state-of-the-art classifiers in the literature. One such technique is to use side information features obtained from the DNS query/response that cannot be easily manipulated by the adversary. Although there are past research works that use DNS features for a retrospective analysis of DNS traffic, to the best of our knowledge, there are no studies that leverage such data for inline detection of DGA domains. In our work, we train machine learning models based on tree ensembles and deep learning for DGA detection using side information (in addition to the domain name), which can be easily obtained in practice without relying on external data sources such as WHOIS. Besides, we also disregard methods that analyze past DNS data to extract side information features, thereby resulting in a relatively lightweight computation for detecting DGA domains in real-time DNS applications. In the end, we also perform an empirical evaluation by applying the best performing classifiers trained using side information on one day of passive DNS traffic to compare its performance against well known state-of-the-art classifier that relies only on a domain name for DGA detection. Results show that classifiers trained using a combination of lexical and side information features, not only provide high performance but are also more robust to adversarial attacks than the classifiers that rely only on the domain name for inline DGA detection.