ENHANCED GATED RECURRENT UNIT DEEP LEARNING MODEL FOR VEHICULAR NETWORKS ANOMALY-BASED INTRUSION DETECTION

Abstract

ABSTRACT

The growing connection and dynamic communication patterns of vehicle-to-everything (V2X) systems provide serious security concerns for modern vehicular networks. In vehicle contexts, traditional intrusion detection techniques are insufficient for identifying new assaults and adjusting to quickly evolving network topologies. An improved Gated Recurrent Unit (GRU)-based deep learning model for anomaly-based intrusion detection in vehicular networks is presented in this study. Multi-scale convolutional neural networks (CNN) with variable kernel sizes (3, 5, and 7) for temporal pattern recognition, principal component analysis (PCA) for dimensionality reduction while maintaining 95% variance, multi-head attention mechanisms for focused feature engineering, statistical feature extraction for capturing distributional characteristics of traffic patterns, and Adaptive Synthetic Sampling (ADASYN) for handling imbalanced datasets are all integrated into the suggested architecture. The hybrid CNN-GRU model combines CNN's spatial feature extraction capabilities with bidirectional GRU layers to capture temporal dependencies. The findings of performance evaluation on five publicly accessible datasets (NSL-KDD, CAN_HCRL_OTIDS, Car-Hacking, Network-Traffic, and Road-Traffic) are remarkable, with accuracy rates ranging from 97.09% to 100%. The CAN_HCRL dataset exhibits perfect classification. The model successfully detects a variety of attack types, such as denial-of-service, replay attacks, and fake information injection, while maintaining low false positive rates (lowest value, 0.01%) and high ROC-AUC scores (0.9000 to 1.0000). ADASYN greatly increased minority class detection by improving class balance by up to 99.4%. By offering a reliable, real-time intrusion detection solution appropriate for safety-critical autonomous driving applications, this research increases vehicular network security.