Recently, reconstruction-based anomaly detection was proposed as an effective technique to detect attacks in dynamic industrial control networks. Unlike classical network anomaly detectors that observe the network traffic, reconstruction-based detectors operate on the measured sensor data, leveraging physical process models learned a priori. In this work, we investigate different approaches to evade prior-work reconstruction-based anomaly detectors by manipulating sensor data so that the attack is concealed. We find that replay attacks (commonly assumed to be very strong) show bad performance (i.e., increasing the number of alarms) if the attacker is constrained to manipulate less than 95% of all features in the system, as hidden correlations between the features are not replicated well. To address this, we propose two novel attacks that manipulate a subset of the sensor readings, leveraging learned physical constraints of the system. Our attacks feature two different attacker models: A white box attacker, which uses an optimization approach with a detection oracle, and a black box attacker, which uses an autoencoder to translate anomalous data into normal data. We evaluate our implementation on two different datasets from the water distribution domain, showing that the detector's Recall drops from 0.68 to 0.12 by manipulating 4 sensors out of 82 in WADI dataset. In addition, we show that our black box attacks are transferable to different detectors: They work against autoencoder-, LSTM-, and CNN-based detectors. Finally, we implement and demonstrate our attacks on a real industrial testbed to demonstrate their feasibility in real-time.@en

In recent years, several schemes have been proposed to detect anomalies and attacks on Cyber-Physical Systems (CPSs) such as Industrial Control Systems (ICSs). Based on the analysis of sensor data, unexpected or malicious behavior is detected. Those schemes often rely on (implicit) assumptions on temporally stable sensor data distributions and invariants between process values. Unfortunately, the proposed schemes often perform not optimally with Recall scores lower than 70% (e.g., missing 3 alarms every 10 anomalies) for some ICS datasets, with unclear root issues. In this work, we propose a general framework to check whether a given ICS dataset has specific properties (stable sensor distributions in normal operations, potentially state-dependent), which then allows to determine whether certain Anomaly Detection approaches can be expected to perform well. We apply our framework to three datasets showing that the behavior of actuators and sensors are very different between Training set and Test set. In addition, we present high-level guides to consider when designing an Anomaly Detection System. @en