Anomaly detection using improved deep SVDD model with data structure preservation

Abstract

• Typical SVDD structures are investigated in view of model depth profiles. • An improved deep SVDD model is proposed by considering the data structure preservation. • The experiments on the benchmark datasets verify the effectiveness of the proposed method. Support vector data description (SVDD) is a classical anomaly detection algorithm. How to develop a deep version of SVDD is one valuable problem in the anomaly detection field. Aiming at this problem, an improved SVDD model called deep structure preservation SVDD (DSPSVDD) is proposed by integrating the deep feature extraction with the data structure preservation. Firstly, the typical SVDD methods are revisited in view of model depth profiles and the limitations of the present deep SVDD model are analyzed. Then in order to extract the deep data features more effectively, an enhanced comprehensive optimization objective is designed for the deep SVDD model by considering both the hypersphere volume minimization and the network reconstruction error minimization simultaneously. The experimental results on the MNIST, Fashion-MNIST, and MVTec AD image benchmark datasets show that the proposed DSPSVDD method achieves the better anomaly detection performance compared with the traditional deep SVDD method.