LCSMC-Net: Lightweight CAN Intrusion Detection via Separable Multiscale Convolution and Attention.

Area of Science:

• Automotive Cybersecurity
• Embedded Systems AI
• Deep Learning for Network Security

Background:

• Controller Area Network (CAN) protocol lacks inherent security, making vehicles vulnerable to cyberattacks.
• Existing deep learning intrusion detection systems are too computationally intensive for automotive microcontrollers.
• There is a critical need for efficient, embedded security solutions for in-vehicle networks.

Purpose of the Study:

• To propose LCSMC-Net, an ultra-lightweight neural network architecture for intrusion detection in resource-constrained CAN environments.
• To develop a practical and efficient edge AI solution for automotive cybersecurity.
• To address the limitations of current deep learning models in embedded automotive systems.

Main Methods:

• Developed Separable Multiscale Convolution Lite (SMC-Lite) blocks for efficient feature extraction.
• Implemented Lightweight Channel-Temporal Attention (LCTA) with linear complexity for adaptive pruning.
• Utilized 6-dimensional CAN-optimized features for aggressive data compression.
• Employed Bayesian hyperparameter optimization and knowledge distillation for model compression.

Main Results:

• LCSMC-Net achieved 99.89% accuracy on CAN and CAN-FD datasets.
• The model has only 9401 parameters and 2.84M FLOPs, significantly reducing computational load.
• Outperformed existing intrusion detection solutions in terms of efficiency and accuracy.
• Demonstrated feasibility for real-time deployment on automotive-grade microcontrollers.

Conclusions:

• LCSMC-Net provides a viable edge AI solution for securing CAN networks against intrusions.
• The proposed architecture meets the stringent real-time and resource constraints of automotive embedded systems.
• This research enables practical deployment of advanced AI-driven security in modern vehicles.