Real-time behavioral DGA detection through machine learning

Abstract

During the last years, the use of Domain Generation Algorithms (DGAs) has increased with the aim of improving the resiliency of communication between bots and Command and Control (C&C) infrastructure. In this paper, we report on an effective DGA-detection algorithm based on a single network monitoring. The first step of the proposed method is the detection of a bot looking for the C&C and thus querying many automatically generated domains. The second phase consists on the analysis of the resolved DNS requests in the same time interval. The linguistic and semantic features of the collected unresolved and resolved domains are then extracted in order to cluster them and identify the specific bot. Finally, clusters are analyzed in order to reduce false positives. The proposed solution has been evaluated over (1) an ad-hoc network where several known DGAs were injected and (2) the LAN of a company. In the first experiment, we deployed different families of malware employing several DGAs: all the malicious variants were detected by the proposed algorithm. In the real case scenario, the algorithm discovered an infected host in a 15-day-long experimental session, while producing a low false-positive rate during the same period.