Automatic Construction of Statechart-Based Anomaly Detection Models for Multi-Threaded Industrial Control Systems

Traffic of Industrial Control System (ICS) between the Human Machine Interface (HMI) and the Programmable Logic Controller (PLC) is known to be highly periodic. However, it is sometimes multiplexed, due to asynchronous scheduling. Modeling the network traffic patterns of multiplexed ICS streams using Deterministic Finite Automata (DFA) for anomaly detection typically produces a very large DFA and a high false-alarm rate. In this article, we introduce a new modeling approach that addresses this gap. Our Statechart DFA modeling includes multiple DFAs, one per cyclic pattern, together with a DFA-selector that de-multiplexes the incoming traffic into sub-channels and sends them to their respective DFAs. We demonstrate how to automatically construct the statechart from a captured traffic stream. Our unsupervised learning algorithms first build a Discrete-Time Markov Chain (DTMC) from the stream. Next, we split the symbols into sets, one per multiplexed cycle, based on symbol frequencies and node degrees in the DTMC graph. Then, we create a sub-graph for each cycle and extract Euler cycles for each sub-graph. The final statechart is comprised of one DFA per Euler cycle. The algorithms allow for non-unique symbols, which appear in more than one cycle, and also for symbols that appear more than once in a cycle.

We evaluated our solution on traces from a production ICS using the Siemens S7-0x72 protocol. We also stress-tested our algorithms on a collection of synthetically-generated traces that simulated multiplexed ICS traces with varying levels of symbol uniqueness and time overlap. The algorithms were able to split the symbols into sets with 99.6% accuracy. The resulting statechart modeled the traces with a median false-alarm rate of as low as 0.483%. In all but the most extreme scenarios, the Statechart model drastically reduced both the false-alarm rate and the learned model size in comparison with the naive single-DFA model.