Abstract
Active learning has been utilized as an efficient tool in building anomalydetection models by leveraging expert feedback. In an active learningframework, a model queries samples to be labeled by experts and re-trains themodel with the labeled data samples. It unburdens in obtaining annotateddatasets while improving anomaly detection performance. However, most of theexisting studies focus on helping experts identify as many abnormal datasamples as possible, which is a sub-optimal approach for one-classclassification-based deep anomaly detection. In this paper, we tackle twoessential problems of active learning for Deep SVDD: query strategy andsemi-supervised learning method. First, rather than solely identifyinganomalies, our query strategy selects uncertain samples according to anadaptive boundary. Second, we apply noise contrastive estimation in training aone-class classification model to incorporate both labeled normal and abnormaldata effectively. We analyze that the proposed query strategy andsemi-supervised loss individually improve an active learning process of anomalydetection and further improve when combined together on seven anomaly detectiondatasets.