Real-time out-of-distribution detection in cyber-physical systems with learning-enabled components

Abstract

Learning-enabled components (LECs) such as deep neural networks are used increasingly in cyber-physical systems (CPS) since they can handle the uncertainty and variability of the environment and increase the level of autonomy. LECs, however, may compromise system safety since their predictions may have large errors, for example, when the data available at runtime are different than the data used for training. This study considers the problem of efficient and robust out-of-distribution detection for learning-enabled CPS. Out-of-distribution detection using a single input example is typically not robust and may result in a large number of false alarms. The proposed approach utilises neural network architectures that are used to compute efficiently the nonconformity of new inputs relative to the training data. Specifically, variational autoencoder and deep support vector data description networks are used to learn models for the real-time detection of out-of-distribution high-dimensional inputs. Robustness can be improved by incorporating saliency maps that identify parts of the input contributing most to the LEC predictions. We demonstrate the approach using simulation case studies of an advanced emergency braking system and a self-driving end-to-end controller, as well as a real-world data set for autonomous driving. The experimental results show a small detection delay with a very small number of false alarms while the execution time is comparable to the execution time of the original LECs.