Stealthy False Data Injection Attack Detection and Localization Using Reduced Sparse Transformer Neural Network

Abstract Cyber-physical power systems have seen a considerable rise in malicious false data injection (FDI) attacks over the last decade. Metering infrastructure is most vulnerable to such attacks as they are spread out over a large topological area, and their location often is accessible to consumers/generators. We consider such a scenario in our study where low magnitude stealthy FDI attacks are injected at different buses of an IEEE 14-bus, 30-bus, and 118-bus systems. We propose a novel reduced sparse transformer (RST) neural network to detect the presence of FDI attack at multiple buses. The proposed RST uses a time series input of past measurements of active power at each bus received from the metering units and uses an encoder-only architecture to predict the presence of an attack at selected buses. We compare results with a baseline softmax or vanilla transformer neural network (TNN), sparsemax attention-based TNN, convolutional neural network long short-term memory (CNN-LSTM), and bidirectional LSTM (bi-LSTM) networks which are state-of-the art recurrent architectures for time series, text, and natural language prediction. The proposed RST architecture shows significant improvement in classification metrics for multi-label and single label cases for each of the attacked bus locations. The GitHub repository can be found here: GitHub Repository.