Data-Driven Measurement Tampering Detection Considering Spatial-Temporal Correlations

摘要

Cyber-attackers could stealthily interfere with the normal operation of power systems by maliciously tampering the measurements transmitted from field devices to control centers. Although some data-driven methods capable of identifying abnormal measurements have been proposed, there are some remaining shortcomings such as low accuracy and slow convergence rate. This paper devises a novel framework by combining convolutional neural networks (CNN) and long-short-term memory (LSTM) networks for detecting tampering attacks based on recognizing spatial-temporal correlations between measurements. Besides, optimization such as the attention mechanism, Dropout layers and SVM are applied to improve the performance of the proposed framework. This paper also introduces how to implement the proposed framework in practical cyber-physical systems and expounds the synergistic relationships between the data-driven detector and the traditional bad data detection module. Case studies prove that the proposed framework has better learning performance than existing ones.