Feature Collusion Attack on PMU Data-driven Event Classification

Event classification is a critical task to ensure the reliability of the power system. Recent developments in event classification methods leverage data-driven techniques with fine-grained Phasor Measurement Units (PMU) data. However, the existing event classification methods are vulnerable to different types of adversarial attacks that can significantly degrade the event classification performance. In this paper, we evaluate the vulnerability of the classification models against feature collision attacks on PMU data. Feature collusion attack leverages a surrogate model to learn the victim’s classification model which in turn makes it a plausible attack strategy for both black-box and white-box settings. Specifically, this attack strategy undermines the accuracy of the classification models by crafting poisonous samples that share common features with benign samples which in turn changes the decision boundaries of the classification models. The experimental results on real-world PMU data in a black-box setting show that generating and adding poisonous samples into the model training dataset can significantly degrade the accuracy of current event classification methods.