Abstract
Intelligent machine learning approaches are finding active use for eventdetection and identification that allow real-time situational awareness. Yet,such machine learning algorithms have been shown to be susceptible toadversarial attacks on the incoming telemetry data. This paper considers aphysics-based modal decomposition method to extract features for eventclassification and focuses on interpretable classifiers including logisticregression and gradient boosting to distinguish two types of events: load lossand generation loss. The resulting classifiers are then tested against anadversarial algorithm to evaluate their robustness. The adversarial attack istested in two settings: the white box setting, wherein the attacker knowsexactly the classification model; and the gray box setting, wherein theattacker has access to historical data from the same network as was used totrain the classifier, but does not know the classification model. Thoroughexperiments on the synthetic South Carolina 500-bus system highlight that arelatively simpler model such as logistic regression is more susceptible toadversarial attacks than gradient boosting.