FACER: A Universal Framework for Detecting Anomalous Operation of Deep Neural Networks

摘要

The detection of anomalies during the operation of deep neural networks (DNNs) is of essential importance in safety-critical applications, such as autonomous vehicles. In the field, classifiers may face rare environmental conditions, unknown objects, hardware failures, and other types of anomalies. Nevertheless, DNNs still predict arbitrarily high class probabilities in these cases and are unable to recognize out of-distribution operation modes. In this paper, we introduce FACER, an efficient and versatile framework that is trainable to detect various types of anomalies in pre-trained DNNs. FACER operates on compressed intermediate feature representations of the supervised network that can be easily obtained. We evaluate the detection of different input corruptions as well as outliers drawn from out-of-distribution datasets with CIFAR10, CIFAR-100 and SVHN classification models. The detection performance of our method is on par with other state-of-the art methods, while our method can be easier implemented and integrated into resource-constrained hardware systems.