One-class Anomaly Detection with Redundancy Reduction and Momentum Mechanism

Abstract

The objective of anomaly detection is to identify the sample which differs in some known data. In practice, anomaly class is usually hard to obtain and consumptive to label, while unsupervised learning and one-class classification are most widely used to solve this problem. Only a set of data from the specific class are given in the training phase, and the remaining categories will be considered as abnormal. In this paper, inspired by the success of deep learning and Support Vector Data Description (SVDD) of decision boundary-based, a novel idea that combining SVDD with redundant information reduction and momentum update mechanism named RRM-SVDD is proposed to address the anomaly detection problem. With the existence of trivial solutions for SVDD, an objective function is designed to avoid such situation by computing the dimension correlation matrix of the output vector from the feature extraction network, while optimizing it as the identity matrix to make any two dimensions as linearly independent as possible in the pretraining phase, that causes the effective for SVDD to describe the distribution of normal data in the feature space and reduce the probability of model collapse. Meanwhile, the momentum update mechanism is applied to learn the global hyperparameter center C by considering the previous epoch information in the next training period. To evaluate the performance of RRM-SVDD, related experiments on MNIST and CIFAR-10 image benchmark dataset have been conducted, achieved state-of-the-art anomaly detection accuracy and robustness in most categories than comparison methods.