Improving Autonomous Vehicle Safety with DF-GRU and VAE-Powered Anomaly Detection

Abstract

As autonomous vehicles (AVs) have become more widespread, there is a mightier necessity for upgraded security actions to assure credible and secure operation in challenging environments. Guaranteeing AV safety needs anomaly detection, which helps find irregularities in sensor data, system performance, and vehicle behavior. In an effort to strengthen AV safety, this work introduces an anomaly detection approach based on a Convolutional Dual Flow Gated Recurrent Unit (DF-GRU) and Variational Autoencoder (VAE) model. The first convolutional layers emphasize collecting spatial attributes and highlighting semantic elements, while the DF-GRU is conscientious for extracting the temporal context from multivariate time series data. Additionally, the presented construction integrates a VAE to detect anomalies. The VAE generates a threshold setting mode to develop anomaly recognition performance. Anomalies are discovered by paralleling the log-likelihood score and the reconstruction loss of the output, with a pre-established threshold value. Estimate the proposed method’s competence using a range of metrics, such as accuracy, precision, recall, F1 score, and a loss. The experiment results show that the recommended system delivers better than the striving procedures. Ultimately, the model's output proves to have the capacity to successfully spot and avoid cyberattacks on AV.