FADEG: Feature-Adaptive Asymmetric Dynamic Reweighting GNN Ensemble-Based Intrusion Detection for Power Systems

Abstract

The cyber-physical integration of power grids creates vulnerabilities to intrusion attacks. However, the existing neural network-based intrusion detection frameworks exhibit limited robustness against malicious threats. Furthermore, these frameworks struggle to effectively manage the pronounced class imbalance inherent in cybersecurity datasets. In response to these challenges, this paper presents a novel feature-adaptive asymmetric dynamic reweighting ensemble graph neural network for network intrusion detection systems in power grids. Our methodology introduces several novel contributions: first adaptive graph construction techniques that provide stronger connectivity for underrepresented classes, second asymmetric model capacity allocation that dedicates more representational power to challenging classifications, and third dynamic loss weighting strategies that evolve during the training process. The proposed framework incorporates an ensemble of diverse graph neural network models with varying architectures. This ensemble significantly improves the detection of normal network traffic while maintaining high accuracy for attack identification. Experimental results demonstrate that our approach achieves more balanced per-class performance compared to traditional models, with a 15%–20% improvement in minority class detection accuracy. Synthetic minority oversampling technique oversampling, specialized interaction features, and nonlinear transformations were implemented to further enhance model robustness. The ensemble’s voting mechanism provides additional resilience against misclassification. This makes the approach particularly valuable for real-world intrusion detection systems that must maintain high detection rates across all traffic types.