Understanding Credibility of Adversarial Examples against Smart Grid

Stability assessment is an important task for maintaining reliable operations of power grids. With increased system complexity, deep learning-based stability assessment approaches are promising to address the shortfalls of the traditional time-domain simulation-based approaches. However, in the field of computer vision, the deep learning models are shown vulnerable to adversarial examples. Although this vulnerability has been noticed by the energy informatics research, the domain-specific analysis on the requirements imposed for implementing effective adversarial examples is still lacking. These attack requirements, albeit reasonable in computer vision tasks, can be too stringent in the context of power grids. In this paper, we systematically investigate the requirements and discuss the credibility of six representative adversarial example attacks for a case study of voltage stability assessment for the New England 10-machine 39-bus system. We show that (1) compromising the voltage traces of half of transmission system buses is a rule of thumb requirement; (2) the universal adversarial perturbations that are independent of the original clean voltage trajectory have the same credibility as the widely studied false data injection attacks on power grid state estimation, while other adversarial example attacks are less credible; (3) the universal perturbations can be effectively defended with strong adversarial training.