One-Class Classification Using Generative Adversarial Networks

Abstract

One-class classification (OCC) problem has drawn increasing attention in recent years. It expands the range of classification from pre-defined categories to undefined categories. Since the vast diversity of negative samples, it is hard to acquire complete knowledge of unknown classes and construct a negative set for training a one-class classifier, which remains a difficult problem. In this paper, we propose a new OCC model by modifying the generative adversarial network model to address the OCC problem. Taking the generator's outputs as outliers, the discriminator in our model is trained with these synthetic data and target training data, and it manages to distinguish them from each other. Moreover, a new evaluation protocol named classification recall index is put forward to indicate the classifier's performances on both positive and negative sets. The extensive experiments on the MNIST dataset and the Street View House Numbers (SVHN) dataset demonstrate that the proposed model is competitive over a variety of OCC methods.