Detection of False Data Injection Attacks in Cyber-Physical Power Systems: An Adaptive Adversarial Dual Autoencoder With Graph Representation Learning Approach

False data injection attacks (FDIAs) are an important network attack threatening the security of power systems to tamper with instruments and measurements. Conventional FDIAs detection approaches are limited to processing the high-dimensional non-Euclidean correlation of grid data. Inspired by the recent advances in deep learning, we propose a novel unsupervised method for FDIAs detection by combining the complementary strengths of dual graph-convolutional autoencoder (DAE) and generative adversarial network (GAN). The technique first proposes the design of adversarial DAE to represent the data dimension reasonably in a low dimension. Among them, the application of GAN establishes not only practical constraints for reconstruction but also adds new strong support for the score calculation of the detection model. Second, the proposal of the graph convolutional strengthens the reasonable representation of the non-Euclidean data of the power system. Finally, considering the nonstationarity of power system performance, we use dynamic thresholds to adaptively fit the detection scores obtained by the model to improve the overall performance of the model comprehensively. We verify the effectiveness of our proposed unsupervised algorithm by performing on IEEE 14-bus and IEEE 118-bus systems. Furthermore, robustness tests in different noise environments demonstrate the excellent generality of the algorithm.