Search for a command to run...
Off-the-shelf intrusion detection systems prove an ill fit for protecting industrial control systems, as they do not take their process semantics into account. Specifically, current systems fail to detect recent process control attacks that manifest as unauthorized changes to the configuration of a plant's programmable logic controllers (PLCs). In this work we present a detector that continuously tracks updates to corresponding process variables to then derive variable-specific prediction models as the basis for assessing future activity. Taking a specification-agnostic approach, we passively monitor plant activity by extracting variable updates from the devices' network communication. We evaluate the capabilities of our detection approach with traffic recorded at two operational water treatment plants serving a total of about one million people in two urban areas. We show that the proposed approach can detect direct attacks on process control, and we further explore its potential to identify more sophisticated indirect attacks on field device measurements as well.