Graphon Neural Networks-Based Detection of False Data Injection Attacks in Dynamic Spatio-Temporal Power Systems

Abstract

Cyberattacks on power systems have doubled due to digitization, impacting healthcare, social, and economic sectors. False data injection attacks (FDIAs) are a significant threat, allowing attackers to manipulate power measurements and transfer malicious data to control centers. In this paper, we propose the use of graphon neural networks (WNNs) for detecting various FDIAs. Unlike existing graph neural network (GNN)-based detectors, WNNs are efficient as they make use of the non-parametric graph processing method known as graphon, which is a limiting object of a sequence of dense graphs, whose family members share similar characteristics. This allows to leverage the learning by transference on the graphs to address the computational complexity and environmental concerns of training on large-scale systems, and the dynamicity resulting from the spatio-temporal evolution of power systems. Through experimental simulations, we show that WNN significantly improves FDIAs detection, training time, and real-time decision making under topological reconfigurations and growing system size with generalization and scalability benefits compared to conventional GNNs.