Enhancing Detection of False Data Injection Attacks in Smart Grid Using Spectral Graph Neural Network

Abstract

The smart grid (SG) exemplifies the utilization of industrial cyber physical systems within the electric power industry. Ensuring information security is an paramount concern in SG. However, false data injection attack (FDIA) poses considerable risk in manipulating data and compromising SG functions. The existing methods that utilize spectral relationships to detect FDIA primarily target sudden changes and cannot be applied to comb-shaped signal variations. So, addressing this issue, this article introduces a spectral graph neural network-based approach utilizing Bernstein polynomials to approximate spectral graph filters for detecting FDIA. The filter coefficients, obtained through neural network training, enabling to create comb-shaped and high-pass spectral filters applied to the different signal variations. To assess the efficacy of our model, we contrast it with other latest methods and conduct experiments on IEEE 14, 30, and 118 bus systems. The results show an average performance improvement 9.22% of our model relative to the latest models.