Abstract
Validation of modern automotive systems like ADAS (Advanced Driver-Assistance Systems) or autonomous driving systems has become a hard-to-crack nut due to the complexity of the system and huge data volumes generated during the testing phase. The classical methods that borrow heavily from predefined rules and thresholds have a hard time spotting the subtle, unexpected, or even emerging anomalous behaviors. A new framework that switches to unsupervised deep learning for anomaly detection is proposed in this paper to automate the entire validation process. By processing the data from in-vehicle networks (such as CAN, LIN, or Ethernet) in the form of multivariate time series, our proposal builds an LSTM-based autoencoder model that equips itself with the nominal system behavior without the need for labeled fault data.