A Generalizable Deep Neural Network Method for Detecting Attacks in Industrial Cyber-Physical Systems

Today's power systems utilize smart technology to improve the efficacy of power distribution. Using cyber-physical components in the power system such as smart grids can introduce vulnerabilities such as false data injection (FDI) that can cost millions. Deep learning (DL) is an emerging technology that mimics the human brain to process complex problems. In DL, relevant features are extracted automatically to make a meaningful decision out of the data. This article proposes an attack detection method that utilizes DL techniques for detecting FDI attacks. The proposed methodology assumed the problem of varying sparsity attacks in the system, in which attacks can occur at any subset of measurements, as well as the problem of imbalanced training data in real systems. Thus, a deep neural network with regularization techniques including dropout layers and adaptive optimization is proposed for superior generalization to varying sparsity in FDI attacks. An experimental environment is established on simulated power systems of varied sizes and compared with alternative state-of-the-art schemes. The proposed scheme outperforms all of them including robustness to data imbalance and also, it takes lesser time than neural networks of the similar architecture.