MCVAE: Multi-channel Variational Autoencoder for Anomaly Detection

Abstract

Unsupervised anomaly detection is a very important problem due to its wide applications in many systems such as the network communication system, the Internet of Things, and the ICS system. Interpretable modeling of heterogeneous data channels is also essential in anomaly detection, due to the intrinsic multi-modality in multi-channel multi-dimension data. Some existing works use variational autoencoder (VAE) for anomaly detection with a single Gaussian distribution model. However, these VAE can not learn the complex distribution between features, and hence cannot make accurate detections. To tackle this challenge, in this paper, we propose a multi-channel VAE model to jointly account for latent relationships across multiple channels. Moreover, we explore the training loss function combining the reconstruction loss of each channel. We also explore the test loss combining the reconstruction probability and reconstruction ratio between global channel information and each channel information. The proposed detector reports an anomaly when the test loss is below a certain threshold. We conduct extensive simulations on a real world dataset and find that our proposed scheme outperforms the state-of-the-art anomaly detection schemes compared with existing methods.