What question did this study set out to answer?

The aim is to develop a robust intrusion detection framework that enhances accuracy while maintaining data privacy in Internet of Vehicles.

April 25, 20260 citationsOpen Access

MSDA-GDS: A Dual-Branch Hybrid Federated Explainable Deep Learning Framework for CAN Bus Intrusion Detection in Internet of Vehicles.

MSMoh’D Suliman Shakkah BABelal Al-sellami AHAbdulnaser A Hagar

Key Points

The aim is to develop a robust intrusion detection framework that enhances accuracy while maintaining data privacy in Internet of Vehicles.
Proposed a dual-branch hybrid federated explainable framework called MSDA-GDS.
Integrated Multi-Scale Dilated Attention and Gated Depthwise Separable branches with Apache Spark preprocessing.
Evaluated on CICIoV2024 with 1,408,219 CAN frames and CIC-IDS-2017 with 2.83 million flows.
Achieved 99.99% accuracy on CICIoV2024 and 99.40% on CIC-IDS-2017.
Federated variant reached 99.97% accuracy under full privacy protection.
Ablation analysis showed gating mechanism and engineered features had notable impacts (∆F1 = −0.21 and ∆F1 = −0.27 respectively).

Abstract

The Controller Area Network (CAN) bus remains critically vulnerable to cyberattacks due to its lack ofauthentication and encryption. Existing intrusion detection systems (IDS) for Internet of Vehicles (IoV) suffer from single-branch architectures that fail to capture multi-scale CAN byte dependencies, centralized training paradigms that compromise vehicular data privacy, and insufficient model interpretability. This paper proposes MSDA-GDS, a dualbranch hybrid federated explainable framework comprising a Multi-Scale Dilated Attention (MSDA) branch with parallel dilated convolutions and channel-spatial attention, and a Gated Depthwise Separable (GDS) branch with learnable gating mechanisms and residual connections, fused via learned attention weighting. The framework integrates Apache Spark-accelerated preprocessing, FedProx federated learning with differential privacy, and multi-method explainability (SHAP, LIME, gradient saliency). Evaluation on CICIoV2024 (1,408,219 CAN frames) and CIC-IDS-2017 (2.83M flows) demonstrates 99.99% and 99.40% accuracy respectively, with the federated variant achieving 99.97% under full privacy protection. Ablation analysis confirms the gating mechanism (∆F1 = −0.21) and engineered features (∆F1 = −0.27) as the most impactful components, while XAI analysis identifies DATA 2, DATA 1, and DATA 3 as the most discriminative byte positions with high cross-method consistency (ρ = 0.978).

AI에게 질문

Bookmark

View Full Paper