Sprecher
Beschreibung
Industrial intrusion detectors are usually tested one input at a time, even though real attacks unfold as campaigns: attackers learn from earlier probes, detectors combine evidence across sensors, and analysts act on the resulting alerts. My dissertation follows this chain.
KEYSER is a 16-phase adversarial machine learning test framework that lets later attack steps reuse knowledge gathered earlier. I evaluated it against two anomaly detectors. KISTE is a multi-stage deep autoencoder for factory networks; KEYSER targets its network-flow stage. KitNET is a lightweight autoencoder-based benchmark operating on the same flow features. Across 30 paired campaigns per detector, reusing that knowledge raised the evasion rate, meaning the share of attack inputs accepted as normal, from 1.3% to 45.4% against KISTE and from 47.7% to 63.1% against KitNET. When I disabled those information links but kept the same schedule, evasion stayed near baseline.
