Design and Evaluation of Artificial Intelligence-Based Intrusion Detection for Smart Grids

Abduljabbar J. Ajeel; Jamal Kh-Madhloom

doi:10.32792/universityofthi-qar.v21i2.498

Authors

Abduljabbar J. Ajeel 2College of Education for Pure Sciences, University of Wasit, 52001, Iraq
Jamal Kh-Madhloom College of Education for Pure Sciences, University of Wasit, 52001, Iraq

DOI:

https://doi.org/10.32792/universityofthi-qar.v21i2.498

Keywords:

Smart Grid Security, Intrusion Detection, CNN, GATv2, Adversarial Training, XAI, Sherlock Dataset

Abstract

Traditional Intrusion Detection Systems (IDS) in smart grids suffer from intrinsic limitations due to their reliance on fixed rules or known attack signatures, rendering them ineffective against unknown threats and zero-day attacks. This paper proposes a novel hybrid deep learning framework that combines Convolutional Neural Networks (CNN) and Graph Attention Networks version 2 (GATv2) to effectively capture local spatial patterns and structural topological relationships in smart grid traffic. Furthermore, the framework enhances system robustness against adversarial attacks using Adversarial Training strategies (FGSM and PGD), and integrates Explainable AI (XAI) techniques, specifically SHAP, to increase transparency. Evaluated on Sherlock, UNSW-NB15, and NSL-KDD datasets, the proposed model achieved superior accuracy (99.92% on Sherlock) and strong robustness against adversarial attacks. Key contributions include a noise-resistant hybrid architecture, a validated robustness framework, and the integration of XAI for operator trust.