Semi-Supervised Adversarial Learning Framework for Controller Area Network Bus Intrusion Detection

Area of Science:

• Cybersecurity
• Machine Learning
• Automotive Engineering

Background:

• Controller Area Network (CAN) buses are crucial for in-vehicle communication but lack security features like authentication and encryption.
• Vulnerabilities in CAN buses allow for malicious attacks, compromising vehicle safety and data integrity.
• Existing deep learning intrusion detection systems (IDS) face limitations with supervised methods requiring extensive labeled data and unsupervised methods yielding high false positives.

Purpose of the Study:

• To propose a novel semi-supervised generative adversarial network (SGAN) framework for robust intrusion detection on CAN bus systems.
• To address the limitations of existing IDSs by combining image-based CAN representation with adversarial learning for improved accuracy and efficiency.
• To develop a practical cybersecurity solution for protecting safety-critical vehicular sensing and control functions.

Main Methods:

• Consecutive CAN messages are transformed into 64x9 grayscale images for analysis.
• A three-phase training approach is employed: initial discriminator training on labeled data, refinement using distribution-level objectives and generated samples, and generator training to produce realistic adversarial samples.
• The framework utilizes a semi-supervised generative adversarial network (SGAN) integrating image-based CAN representation and adversarial learning.

Main Results:

• Achieved an average accuracy of 99.73% and an F1-score of 99.63% on detecting unknown attacks using leave-one-class-out experiments on the HCRL car-hacking dataset.
• Demonstrated high performance in identifying novel and unseen cyber threats within the CAN bus.
• The model is computationally efficient with only 0.21 million parameters and 3.25 million floating-point operations (FLOPs), making it suitable for resource-constrained automotive platforms.

Conclusions:

• The proposed SGAN framework offers a practical and effective solution for enhancing cybersecurity in connected vehicles by detecting CAN bus intrusions.
• The method successfully overcomes the trade-offs associated with traditional supervised and unsupervised deep learning IDSs.
• The developed intrusion detection system is efficient and accurate, paving the way for enhanced security in automotive sensing and control applications.