Behavioral Profiling of Scada Network Traffic Using Machine Learning Algorithms.

摘要

Mixed traffic networks containing both traditional ICT network traffic and SCADA network traffic are more commonplace now due to the desire for remote control and monitoring of industrial processes. The ability to identify SCADA devices on a mixed traffic network with zero prior knowledge, such as port, protocol or IP address, is desirable since SCADA devices are communicating over corporate networks but typically use non-standard ports and proprietary protocols. Four supervised ML algorithms are tested on a mixed traffic dataset containing 116,527 dataflows from both SCADA and traditional ICT networks: Naive Bayes, NBTree, BayesNet, and J4.8. Using packet timing, packet size and data throughput as traffic behavior categories, this research calculates 24 attributes from each device dataflow. All four algorithms are tested with three attribute subsets: a full set and two reduced attribute subsets. The attributes and ML algorithms chosen for experimentation successfully demonstrate that a TPR of .9935 for SCADA network traffic is feasible on a given network. It also successfully identifies an optimal attribute subset, while maintaining at least a .99 TPR. The optimal attribute subset provides the SCADA network traffic behaviors that most effectively differentiating them from traditional ICT network traffic.