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

Embedding signal processing in the front-end of radiation detectors represents an approach to cope with the growing complexity of nuclear imaging scanners with increasing field of view (i.e., higher number of channels). Machine learning (ML) offers a good compromise between intrinsic image reconstruction performance and computational power. While most hardware accelerators for ML are based on digital circuits and, thus, require the analog-to-digital conversion of all individual signals from photodetectors, an analog approach allows to streamline the pipeline. We present the study of an analog accelerator implementing a neural network (NN) with 42 neurons in a 0.35- μ m CMOS process node. The specific target is the reconstruction of the position of interaction of gamma-rays in the scintillator crystal of Anger cameras used for PET and SPECT. This chip can be used stand-alone or monolithically integrated within the application specific integrated circuit (ASIC) for the filtering of current signals from arrays of silicon photomultipliers (SiPMs). Computation is performed in charge domain by means of crossbar arrays of programmable capacitor. The architecture of the 64-input ASIC and the training of the NN are presented, discussing the impact of weight quantization on 5 bits. From MATLAB and circuit simulations, consistent with ASIC topology and operations, the NN capabilities were tested using two different datasets, obtained from both simulated data and experimental data, both based on PET detector composed by a monolithic scintillator crystal readout by an 8×8 array of SiPMs. Simulations show an achievable spatial resolution better than 2-mm full-width-at-half-maximum with a 10-mm thick crystal, a max. count rate of 200kHz and the energy efficiency per inference is estimated to be of 93.5GOP/J, i.e., competitive with digital counterparts, with an energy consumption of 38nJ per inference and area of 23mm2.@en