Bagging deep autoencoders with dynamic threshold for semi-supervised anomaly detection

Abstract

In the field of anomaly detection, anomalies are usually very rare compared with normal samples, which is not conducive to the construction of anomaly detection model. In this paper, we propose a semi-supervised anomaly detection algorithm based on deep autoencoder. With this algorithm, only normal samples are needed to train anomaly detection model. To improve the robustness of the algorithm, Bagging ensemble method is used to train and combine multiple deep autoencoders. In the process of Bagging, dynamic threshold for anomaly detection is applied to increase the diversity of individual autoencoder. Compared with other semi-supervised methods including one-class SVM, SOM and K-Means, our proposed method has obvious superiority in the behavior of anomaly detection.