Hybrid Features-Based Intrusion Detection for The Internet of Vehicles using Dynamic Adaptation

Abstract

Abstract: The evolving landscape of the Internet of Vehicles (IoV) has brought to the forefront a discernible array of challenges about network security. In response, this study delves into applying deep learning-based intrusion detection techniques to fortify the IoV against potential network threats. Notably, prevailing approaches often rely on a singular deep learning model for either temporal or spatial feature extraction, with a serial sequence of spatial feature extraction followed by temporal feature extraction. Such methodologies tend to exhibit shortcomings in adequately capturing the Spatiotemporal intricacies inherent in IoV dynamics, thereby adversely impacting intrusion detection efficacy, and contributing to an elevated false-positive rate. To address these challenges, this research proposes an innovative intrusion detection method tailored for the IoV, premised on the parallel analysis of Hybrid features. The methodology commences with the construction of an optimal feature subset based on inter-feature correlations within IoV traffic. Subsequently, a parallelized application of Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) architectures is employed to extract spatio-temporal features from IoV traffic. In acknowledgment of the dynamic nature of IoV environments, a novel Dynamic Adaptation Mechanism is introduced. This mechanism continuously monitors the real-time IoV traffic, detecting feature drifts and triggering adaptations in the spatiotemporal feature extraction process. The adaptation requirements are then seamlessly communicated to the intrusion detection model through a Feedback Loop for Model Updating, ensuring that the model remains adept at discerning emerging network threats. The culmination of this process involves the fusion of parallelly extracted spatiotemporal features, facilitated by the self-attention mechanism. Subsequently, intrusion detection is executed utilizing a Multilayer Perceptron, providing a comprehensive framework that dynamically adapts to the evolving IoV environment. Empirical assessments utilizing the NSLKDD dataset demonstrate the efficacy of the proposed method, manifesting in a notable 2.05% reduction in the false-positive rate. Additionally, the proposed method surpasses baseline performance metrics, including accuracy and F1 score, thereby affirming its proficiency in enhancing intrusion detection capabilities within the Internet of Vehicles paradigm, especially in the context of the introduced Dynamic Adaptation Mechanism.