Graph attention and Kolmogorov-Arnold network based smart grids intrusion detection

Area of Science:

• Cybersecurity
• Electrical Engineering
• Computer Science

Background:

• The digital transformation of power systems increases cyberattack risks due to complexity and interconnectivity.
• Traditional intrusion detection systems struggle to integrate physical power grid data and model complex attack patterns.
• Existing graph neural network (GNN) methods often overlook physical device interactions and rely on fixed activation functions.

Purpose of the Study:

• To develop an advanced intrusion detection framework for smart grids that accurately captures intricate device interactions and complex attack patterns.
• To enhance the precision of intrusion detection in critical power infrastructure by integrating physical and network data.
• To overcome the limitations of traditional GNNs in representing nonlinear attack behaviors.

Main Methods:

• Introduced GraphKAN, a novel framework combining Graph Attention Network (GAT) and Kolmogorov-Arnold Network (KAN).
• Constructed a comprehensive graph representation including power, IT, and communication devices, with edges representing physical and logical dependencies.
• Employed GAT with multi-head attention for dynamic node weighting and KAN with learnable B-spline activations for enhanced nonlinear feature expression.

Main Results:

• GraphKAN achieved high detection accuracies: 97.63% (binary), 98.66% (ternary), and 99.04% (37-class).
• Demonstrated significant accuracy improvements over state-of-the-art models (e.g., 5.73% gain over GA-RBF-SVM).
• Validated performance on datasets from Mississippi State University and Oak Ridge National Laboratory.

Conclusions:

• GraphKAN effectively enhances intrusion detection accuracy in smart grids by integrating physical and network information.
• The framework shows robust performance in identifying complex cyberattack patterns.
• The combination of GAT and KAN offers a powerful approach for securing critical power infrastructure.