A Uniform Framework for Anomaly Detection in Deep Neural Networks

Deep neural networks (DNN) can achieve high performance when applied to In-Distribution (ID) data which come from the same distribution as training set. When presented with anomaly inputs not ID, outputs of a DNN should be regarded meaningless. However, modern often predict an ID class confidence, is dangerous and misleading. In this work, we consider three classes inputs, (1) natural different than trained for, known Out-of-Distribution (OOD) samples, (2) crafted generated by attackers, adversarial (AD) (3) noise (NS) samples meaningless data. We propose framework that aims detect all these anomalies for pre-trained DNN. Unlike some existing works, our method does require preprocessing input data, nor it dependent any OOD set or attack algorithm. Through extensive experiments over variety models detection aforementioned anomalies, show in most cases outperforms state-of-the-art methods identifying anomalies.