Generative Abnormal Data Detection for Enhancing Cellular Vehicle-to-Everything Based Road Safety

Under the framework of Cellular-Vehicle-to-Everything (C-V2X) technology, although vehicles can avoid potential risks and improve traffic efficiency, shared vehicle data may have defects or faults due to inevitable environmental noise or potential sensor failures, which could pose dangers to drivers. Therefore, detecting anomalies in data transmitted via C-V2X is crucial, particularly for the driving control messages, i.e., Basic Safety Messages (BSM). However, anomaly detection in BSM data faces multiple challenges. First, BSM data contains rich driving details, necessitating modeling its high variability to better learn complex and nonlinear spatiotemporal relationships. Second, the rarity of anomalous events and the potential diversity of normal behaviors make defining anomalies more complex, increasing the difficulty of anomaly detection. Third, extracting meaningful information from a large amount of data and understanding the abstract patterns or regularities within it can also be challenging for effective reasoning at the data level. To address these challenges, we propose a hybrid generative model named CoGAN, which combines Variational Autoencoder (VAE) and Generative Adversarial Network (GAN) to implicitly learn the feature representation of normal data in an unsupervised manner. Specifically, the VAE is responsible for learning the distribution of normal data, and capturing the fundamental patterns and structures of the data; meanwhile, the discriminator is dedicated to enhancing the model's ability to learn the distribution of normal data, refining the model's understanding of data through the introduction of an adversarial process. CoGAN explores the distribution characteristics of normal vehicle behavior data by jointly learning the generation process and variational inference of BSM data, thereby achieving the purpose of anomaly detection.