Adversarial Attacks in Industrial Control Cyber Physical Systems

Machine learning and deep learning algorithms have been the frontier of artificial intelligence (AI) that reshape the current landscape of computing. However, the trustworthiness and reliability of AI models is a growing concern after they are being widely deployed. Adversarial attacks are cyber-related attacks that target machine learning and deep learning algorithms so that cause the training of the network to be fooled, and inaccurate predictions to be made. As result, adversarial attacks can affect critical AI systems, such as industrial control cyber-physical systems. They rely on machine learning models to perform day to day functions, this alone makes them prone to adversarial attacks since machine learning models are highly vulnerable to adversarial examples. Research on the consequences of these types of attacks can lead to insight on what can be done to prevent these malicious attacks if they were to be imposed on a vulnerable system. In this study, three adversarial cyber-attacks, specifically on power systems, are presented. Specifically, the Fast Gradient Sign Method, DeepFool, and Jacobian-Based Saliency Map Attacks were utilized to generate adversarial examples for machine learning and deep learning. The outcome of this study clearly indicates that adversarial attacks have negative implications on the performance of deep neural networks of cyber physical systems.