A representation learning-based time series label propagation for smart grid attack detection

The rising incidence of cyber attacks targeting smart grids underscores the urgent need for robust protection mechanisms. Most of the existing smart grid cyber attack detection schemes need correctly labeled data for training. However, a scarcity of labeled smart grid attack instances leads to difficulty in learning accurate features in training data. Hence, to improve the robustness of the detection algorithms, this work proposes a graph-based semi-supervised method to classify the unlabeled time series smart grid data. Unlike the traditional machine learning (ML) methods that depend exclusively on labeled data, the proposed method implements a semi-supervised algorithm to reduce the dependency on labels. The proposed method integrates representation learning to identify useful features from the input dataset, thereby enriching the detection model’s capacity to recognize and respond to novel threats. Following this, a semi-supervised algorithm using label propagation for time series is employed to recognize unknown attack instances. Comprehensive evaluations conducted in simulated smart grid environments demonstrate that the proposed method outperforms traditional approaches in detection accuracy, reducing false positives and computational efficiency. The findings of this research illustrate the promise of semi-supervised learning in providing a more adaptive, scalable, and resilient defence against cyber threats in smart grid systems.