BOTection: Bot Detection by Building Markov Chain Models of Bots Network Behavior Bushra A. Alahmadi Enrico Mariconti Riccardo Spolaor University of Oxford, UK University College London, UK University of Oxford, UK
[email protected] [email protected] [email protected] Gianluca Stringhini Ivan Martinovic Boston University, USA University of Oxford, UK
[email protected] [email protected] ABSTRACT through DDoS (e.g. DDoS on Estonia [22]), email spam (e.g. Geodo), Botnets continue to be a threat to organizations, thus various ma- ClickFraud (e.g. ClickBot), and spreading malware (e.g. Zeus). 10,263 chine learning-based botnet detectors have been proposed. How- malware botnet controllers (C&C) were blocked by Spamhaus Mal- ever, the capability of such systems in detecting new or unseen ware Labs in 2018 alone, an 8% increase from the number of botnet 1 botnets is crucial to ensure its robustness against the rapid evo- C&Cs seen in 2017. Cybercriminals are actively monetizing bot- lution of botnets. Moreover, it prolongs the effectiveness of the nets to launch attacks, which are evolving significantly and require system in detecting bots, avoiding frequent and time-consuming more effective detection mechanisms capable of detecting those classifier re-training. We present BOTection, a privacy-preserving which are new or unseen. bot detection system that models the bot network flow behavior Botnets rely heavily on network communications to infect new as a Markov Chain. The Markov Chains state transitions capture victims (propagation), to communicate with the C&C server, or the bots’ network behavior using high-level flow features as states, to perform their operational task (e.g.