A Practical and Stealthy Adversarial Attack for Cyber-Physical Applications

Abstract

Adversarial perturbations on misleading a well-trained machine learning (ML) model have been studied in computer vision (CV) and other related application areas. However, there is very limited focus on studying the impact of adversarial perturbations on ML models used in data-driven cyber-physical systems (CPSs) that normally have complex physical and mechanical constraints. Because of the complex physical and mechanical constraints, called domain-knowledge constraints in our paper, established gradient-based adversarial attack methods are not always practical in CPS applications. In this paper, we propose an innovative CPS-specific adversarial attack method that is able to practically compromise the ML-based decision makings of CPSs while maintaining stealthy by meeting the complex domain-knowledge constraints. In the section of performance evaluations, different scenarios are considered to illustrate the effectiveness of the proposed adversarial attack method in achieving a high success rate as well as sufficient stealthiness in CPS applications.

Tasks

Reproductions