Protecting cyber physical systems using neural networks

Abstract

The versatile, distributed, and heterogeneous nature of Cyber Physical Systems (CPSs) has made it integral to the fourth industry revolution. However, this has also made them prone to various cyber and/or physical security threats and attacks. Anomaly Detection is an effective solution to address these concerns, and one of the approaches involves the use of semi-supervised deep neural networks. Deploying these networks closer to the edge increases privacy, and reduces latency and network load. Therefore, the architectural design of these models should be optimized to cater to the power consumption, memory, and computational constraints of a microcontroller (MCU). This project studies the performance of one-dimensional convolutional neural network (CNN) models, designed for uni-variate time series prediction and anomaly detection, while being constrained for embedding into edge devices. Multiple variants of resource efficient convolutional networks were tested on the Secure Water Treatment (SWaT) dataset. After which, they were compared for their time series prediction accuracy, anomaly detection accuracy, and training time.