"Automatically Generating Models for Botnet Detection"
Friday, July 17. 2009
One of the papers that we will publish at the European Symposium on Research in Computer Security (ESORICS'09) focusses on the problem of detecting bots within a given network. Previous research focussed for example on detecting bots using human-generated signatures and anomaly detectors (e.g., BotHunter) or correlating the activity of individual hosts in order to find machines that react in lockstep (e.g., BotMiner or TAMD). We present a system that automatically generates signatures which encapsulate the behavior of an infected machine. The important observation is that the principle behind bots is that they receive a command from the botherder and then respond in a specific way. Using real-world traces of many botnets we show that it is possible to spot the bot responses in the network traces using a change point detection algorithm. Based on this information we can then identify the commands and we use all information to then encode a signature which we map into Bro rules. Experiments in different networks show that this approach outperforms BotHunter. More information about the approach is available in the paper and all the gory details are published in a technical report. Abstract: A botnet is a network of compromised hosts that is under the control of a single, malicious entity, often called the botmaster. We present a system that aims to detect bots, independent of any prior information about the command and control channels or propagation vectors, and without requiring multiple infections for correlation. Our system relies on detection models that target the characteristic fact that every bot receives commands from the botmaster to which it responds in a specific way. These detection models are generated automatically from network traffic traces recorded from actual bot instances. We have implemented the proposed approach and demonstrate that it can extract effective detection models for a variety of different bot families. These models are precise in describing the activity of bots and raise very few false positives.
This work is a collaboration with Peter Wurzinger, Leyla Bilge, Jan Goebel, Christopher Kruegel, and Engin Kirda. And the word cloud on the top of the posting is generated with the help of http://www.wordle.net/.
