A Hybrid Autoencoder and Density Estimation Model for Anomaly Detection

A novel one-class learning approach is proposed for network anomaly detection based on combining autoencoders and density estimation. An autoencoder attempts to reproduce the input data in the output layer. The smaller hidden layer becomes a bottleneck, forming a compressed representation of the data. It is now proposed to take low density in the hidden layer as indicating an anomaly. We study two possibilities for modelling density: a single Gaussian, and a full kernel density estimation. The methods are tested on the NSL-KDD dataset, and experiments show that the proposed methods out-perform best-known results on three out of four sub-datasets.