| dc.description.abstract | Learning-Enabled Components (LECs) such as deep neural networks are used increasingly in Cyber-Physical Systems (CPSs) since they can handle the uncertainty and variability of the environment and increase the level of autonomy. However, Out-Of-Distribution (OOD) data, which are different than the data used for training, may cause the predictions of LECs to have large errors, and compromise the safety of the overall system. Therefore, detection of OOD data is pivotal to ensure the safe and reliable operation of CPS. Based on the Inductive Conformal Anomaly Detection (ICAD) framework, this dissertation presents several learning-based techniques for efficient and robust detection of OOD data in CPS. First, Variational Autoencoder (VAE) and deep Support Vector Data Description (deep SVDD) networks are used to learn models for the real-time detection of OOD high-dimensional inputs. Second, we discuss the causes of OOD data and define various types of OOD data in learning-enabled CPS. In order to enable the detection to take into consideration both LEC inputs and outputs, a VAE for classification (regression) model is utilized to detect different types of OOD data for classification (regression) problems. Third, an Adversarial Autoencoder (AAE) model is also employed to detect various types of OOD data for classification problems. Last, we propose a novel sequential generative model and utilize it to detect anomalous behavior in high-dimensional time-series data. The experimental results demonstrate that
all the proposed approaches can detect the OOD data with a small number of false alarms, while the approaches are computationally efficient and can be used for online detection. | |
