No abstract

We present a novel semi-supervised framework for training classifiers and simultaneously detecting out-of-distribution inputs. We do this by training on an closed classification dataset and an auxiliary simulated-open dataset, which consists of examples from outside the closed set. Through unsupervised learning and incorporating a class-distance value for each known class, we can identify out-ofdistribution RF devices with state-of-the-art accuracy. We define metrics for quantifying robustness in terms of both classification and Open Set Recognition (OSR). Finally, we discuss uncertainty estimation and calibrate our open set predictions so that they represent confidence. CCS CONCEPTS• Computing methodologies → Anomaly detection.