Problem
A security monitoring sensor on a SPAN (mirror) port at industrial sites sees every node on the network but cannot tell what any of them is. Scanning with Nmap is not allowed: any request to a production network can break something and bring a huge fine. Engineers spend a lot of time on manual node inventory. The sensor is a low-end device with limited CPU and memory, so it can read only packet headers. Users also want anomalies reported.
Example from the article: a node on a plant network starts making many HTTP requests, which it never did before, and the operator must be told that its behavior has changed.
Solution in brief
Build a profile for each node from passively observed traffic: how often it uses each protocol flow and how often it talks to nodes of known types. A naive Bayes classifier trained on profiles from already inventoried sites predicts the device type and operating system, and an engineer confirms it. For anomalies, cluster the profiles hierarchically in each time slice and flag nodes that moved to another cluster.
Questions this case answers
- How to classify entities and detect behavioral changes from passively observed event streams when active probing is forbidden and compute is tight
- How to identify what kind of device each network node is without sending any requests to the network?
- How to detect anomalous behavior of network nodes from packet headers on a weak sensor?
- How to explain to an operator why a node was flagged as anomalous?
Tools and methods
- Flow-based node profiles — vectors of protocol usage frequencies built from packet headers only.
- Naive Bayes classifier — predicts device type and operating system from a profile.
- Hierarchical clustering — run per time slice and compared with the previous slice to find nodes that changed cluster.
- Cophenetic distance — severity score for an anomaly.
Google or a one-time email code. No password.