Deep Learning Algorithm using Transfer-Entropy measures for Anomaly Detection in Cyber-Physical Systems

Abstract

Generally, in cyber-physical systems, there are various attacks detected such as internet-based load altering attacks, False-Data Injection Attack (FDIA), stealthy deception attacks, covert attacks, time synchronization attacks, etc. Over the past decades, attack detection and secure control system design has a high interest due to the rapid growth of cyber security challenges by sophisticated attacks in cyber-physical system like Internet-of-Things (IoT). Among various techniques, Transfer Entropy Measure (TEM) was introduced to detect four types of attacks like Denial-of-Service (DoS), replay, innovation-based deception attack and data injection attacks. Since, it discovers the interaction behavior among pairs of entities generating by each cyber-physical systems. As well, conventional machine-learning based attack detection mechanisms have been successfully employed in IoT i.e., wireless sensors to detect cyber-attacks. However, such mechanisms have less accuracy and scalability with high computational complexity. Hence in this article, a novel distributed deep learning algorithm is proposed for cyber attack detection in IoT since deep learning algorithms try to learn high-level features from data in an incremental manner and solve the problem end to end. Here, the transfer-entropy is measured with different parameters like node, network and channel for sensor measurements. Then, the obtained values are gathered as training dataset. Subsequently, Artificial Neural Network (ANN) and Deep Neural Network (DNN) are trained with training dataset to detect the existence of the attacks in cyber-physical system. Finally, the average detection accuracy values of ANN and DNN are evaluated through the simulation results as 98.9% and 99.6% respectively.