Umeå universitets logga

Modern automotive systems’ increasing complexity and connectivity necessitate robust anomaly detection methods to identify malfunctions, cybersecurity threats, and performance issues. This thesis evaluates multiple unsupervised anomaly detection and classification techniques explicitly tailored for automotive Ethernet environments. Three primary approaches to anomaly detection are examined: a multilayer perceptron (MLP) with an autoencoder architecture, a Gaussian mixture model, and a long short-term memory (LSTM) model with an autoencoder architecture. These approaches are evaluated on combinations of three network interfaces across two datasets from two distinct car models using the same vehicle platform.

Experiments assess these methods’ accuracy, usefulness, and practical feasibility using two proprietary datasets collected with the Secure Gateway team at Volvo Cars, where anomalies are simulated through packet injection and modification. Given the hardware constraints inherent in vehicular environments, computational time is evaluated to determine each method’s suitability. The results offer insights into the most effective techniques for enhancing vehicle safety, reliability, and security, contributing to the advancement of intelligent automotive systems.