Anomalous connection detection using combination of online machine learning models

Abstract

Systems and methods are disclosed to implement a cyberattack detection system that monitors a computer network for suspected lateral movement. In embodiments, the system employs multiple machine learning models to analyze connection data of a network to identify anomalies in the network's connection behavior. The models are updated incrementally using online machine learning methods that can be performed in constant time and memory. In embodiments, the system uses an incremental matrix factorization model and a connection count fitting model to generate anomaly scores for each connection. Connection paths are constructed for acyclic sequences of time-ordered connections observed in the stream. The paths are evaluated based on the anomalies scores of their individual connections. Paths that meet a detection criterion are reported to analysts for further review. Because the detection models are online models, they are continuously updated based on newly observed data, without having to store the new observation data.