Enhancing Autonomous Vehicle Safety via DF-GRU and VAE Based Anomaly Detection

Abstract

The development of autonomous vehicles (AVs) has accelerated the need for advanced safety mechanisms to ensure reliable and secure operation in real-world environments. A crucial aspect of maintaining AV safety is anomaly detection, which helps identify irregularities in vehicle behavior, sensor data, and system performance. This paper introduces a framework designed for detecting anomalies that utilizes a Convolutional Dual Flow Gated Recurrent Unit (DF-GRU) and Variational Autoencoder (VAE) model, designed to enhance AV safety. The initial convolutional layers focus on extracting spatial features and highlighting semantic details, whereas the DF-GRU is responsible for capturing the temporal context from multivariate time series data. Additionally, suggested Framework incorporates a VAE to identify anomalies. A threshold setting strategy is developed in the Variational Autoencoder to enhance the performance of anomaly detection. Anomalies are detected by analyzing the reconstruction loss of the output alongside the log-likelihood score, using a defined threshold value. To evaluate the effectiveness of the proposed framework, use various metrics such as accuracy, precision, recall, F1 score, and an analysis of the loss curve. The experimental outcomes express that the suggested approach excels existing competing algorithms. The model's performance and results in detecting anomalies show that it effectively identifies unusual patterns in multivariate time series data. Ultimately, it is concluded that the model's output can be leveraged to intelligently identify and prevent cyber-attacks on autonomous vehicles.