On the Efficacy of Anomaly Detection in Process Control Networks Alfonso Valdes SRI International alfonso.valdes@sri.com April, 2006
Background Digital automation has made control systems safer, more productive Formerly, purpose-built, isolated, proprietary protocols and platforms Increasingly, commodity platforms and protocols encapsulating legacy, integration to enterprise systems Intelligent end devices with embedded OS and configured over web interface Security practices lag enterprise security Best practice documents emerging Widely distributed systems with weak perimeter control IDS/IPS still relatively novel in PCS Threat not well understood
Critical Need The National Critical Infrastructure needs defenses that detect and prevent cyber and blended cyber/physical attack, enable effective response, and facilitate timely recovery Such defenses must secure the present heterogeneous environment of legacy and modern systems, as well as get and stay ahead of the technology curve
Anomaly Detection Advantage over signature systems: potential to detect unknown attacks Not widely used in enterprise IDS/IPS False alarms Malicious is not always anomalous, anomalous is not always malicious (McHugh) Learning based Statistical N-Grams Specification Based Difficult to specify real systems at adequate fidelity
Hypothesis: AD Will be more Effective in Control Systems Topology is relatively static System mission is relatively narrow in scope Many important messages are regularly timed Both learning and spec based AD may be more feasible and effective Room to explore information theoretic, frequency, wavelet, other novel approaches Counter trend: adoption of sensor nets (large number of nodes, nodes come and go)