Detecting abnormal behavior is an important objective in security monitoring, but is extremely challenging as we mostly are expected to detect "unknown unknowns." We can, however, use an entity's past behavior to measure how much of what we observe today deviates from normal behavior. In this way we can detect unknown, hidden and insider threats early on to stay ahead of advanced threats. This talk presents a unified, scalable framework for anomaly detection that is built on the frequent itemset mining technique. The premise is that if we can align an event with more frequent patterns observed in history, then the event is unlikely to be an anomaly. By mining through an extensive set of features and feature co-occurrences, the model can accurately capture the normal behaviors. Any new behaviors can then be scored. At which point, any new rare co-occurrences of events can be detected and sent to analysts and SOC teams for rapid investigation.

Speaker(s)
Nancy Jin, Data Scientist, Splunk
Ping Jiang, Sr. Software Engineer in Test, Splunk

Slides PDF link - https://conf.splunk.com/files/2019/slides/SEC1230.pdf?podcast=1576909588