Detection of data-driven blind cyber-attacks on smart grid: A deep learning approach

Abstract

An immense challenge to future smart cities is providing resilience against cyber-attacks on critical infrastructures like the smart grid. Data-driven cyber-attack strategies like the false data injection attack (FDIA) can modify the states of the grid, hence posing a critical scenario. With an accurate knowledge of measurement subspace, this work demonstrates an effective blind FDIA formulation strategy. The current trend in the identification of such attacks is generally confined to their presence detection within the measurements, while their intrusion points remain concealed. To alleviate this concern, this study promotes an effective execution of novel robust, nonlinear deep learning models which can not only effectually identify the presence but also the exact locations of intrusions of blind attacks in real-time while working concurrently with the conventional bad data detector, thus furnishing a cost-effective approach. Moreover, such neural network models demonstrate a multilabel classification strategy by capturing the inconsistency with co-occurrence dependency of the attack vectors introduced into the raw measurements. Furthermore, these deep learning structures prove to be model-free, hence attacks can be determined without any statistical knowledge of the grid. The performance of the proposed framework is evaluated utilizing the standard IEEE test bench undertaking diverse noise and attack scenarios. © 2023 Elsevier Ltd