Synthetic false data injection-inspired perturbation framework for power grid attack detection using graph neural networks and deep learning models

Abstract Modern power grids are exposed to False Data Injection (FDI) attacks due to their increasing dependency on digital sensing and measurement infrastructure. The adversaries deliberately manipulate sensor data for compromising control, system monitoring, and overall grid stability. Undetected attacks trigger cascading failures and severe economic disruption and bring challenges for developing a robust, data-driven detection framework for safeguarding critical energy infrastructure. The present study undertakes this challenge by introducing a structured, coordinated, feature-group-based synthetic FDI-inspired perturbation framework. Five architecturally different attack detection models deployed are Multilayer Perceptron (MLP), a Pure LSTM Autoencoder, Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Standard GraphSAGE with an identical 20-step pipeline. A common k-NN cosine-similarity graph (k = 10) is used for the graph-based models. A model-specific hyperparameter grid search is conducted, which comprises 24 configurations for MLP, GCN, GAT, and GraphSAGE and 8 configurations for the LSTM autoencoder, reflecting its distinct unsupervised architecture. A 10-seed robustness test is performed to ensure rigorous, reproducible, and fair cross-model comparison. Under single-run evaluation, the GraphSAGE architecture achieves the strongest overall detection performance (ROC - AUC 0.7277 and PR - AUC 0.4679), outperforming the LSTM Autoencoder (ROC - AUC 0.6540), GCN (ROC - AUC 0.6284), MLP (ROC - AUC 0.5959), and GAT (ROC - AUC 0.4726). In multi-seed robustness evaluation results the GAT architecture emerges as the most reliable and stable detector with an accuracy of 0.8189 ± 0.0123 over ten random initialisations. The findings of this research motivate reporting both single-run and multi-seed robustness metrics when evaluating FDI attack detection models intended for real-world grid deployment. The novelty of this work lies in providing a unified, leakage-free comparative framework spanning graph- and non-graph-based deep learning architectures under identical experimental conditions. As the perturbations are FDI-inspired rather than state-estimation-aware, it provides a realistic and extensible benchmark for future power-grid cybersecurity research.