An intrusion detection model based on Convolutional Kolmogorov-Arnold Networks.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• Artificial Neural Networks (ANNs) are widely used in intrusion detection but suffer from large model sizes and lack of interpretability.
• Existing ANNs present challenges like parameter bloat and opaque decision-making processes in cybersecurity applications.

Purpose of the Study:

• To introduce Convolutional Kolmogorov-Arnold Networks (CKANs) as a novel, interpretable, and accurate intrusion detection model.
• To address the limitations of traditional ANNs in terms of model size and interpretability for intrusion detection.

Main Methods:

• Developed Convolutional Kolmogorov-Arnold Networks (CKANs) based on the Kolmogorov-Arnold representation theorem.
• Integrated attention mechanisms into the CKAN architecture for enhanced computational logic.
• Trained and validated the model using the CICIoT2023 and CICIoMT2024 datasets.

Main Results:

• The CKAN-based intrusion detection model demonstrated high accuracy with significantly fewer parameters compared to other methods.
• Achieved attractive application prospects in intrusion detection due to improved accuracy and reduced model complexity.
• While parameter efficiency is high, the model did not outperform existing methods in memory usage, execution speed, or energy consumption.

Conclusions:

• CKANs present a promising advancement for developing interpretable and accurate intrusion detection systems.
• The proposed model effectively tackles the parameter bloat and interpretability issues inherent in traditional ANNs.
• Further research may be needed to optimize CKANs for efficiency in terms of memory, speed, and energy consumption.