Generative adversarial dimensionality reduction for diagnosing faults and attacks in cyber-physical systems

Abstract

In cyber-physical systems, transforming a large amount of data collected from various sensors onto informative low-dimension data is of paramount importance for efficient monitoring and safe and secure operation of the system. To this aim, this paper proposes two novel dimensionality reduction techniques, where each makes use of two duelling neural networks along with two newly defined constraints of class separability and affinity correlation. Using the original distribution of the high-dimensional data, the goal is to achieve an ideal distribution in a lower-dimensional feature space, while preserving the correlation of features and discriminating samples of distinct classes. These classes are the system states including faults and cyber-attacks. The proposed novel techniques are compared with state-of-the-art dimensionality reduction techniques over several datasets collected from a cyber-physical system. The attained results show that the proposed techniques significantly outperform other techniques.