Anomaly detection based on Deep Autoencoding Gaussian Mixture Model (DAGMM)

Description

According to Anomaly Detection: A Survey, unsupervised anomaly methods can be classified into following categories:

• Classification based techniques
• Clustering based techniques
• Nearest neighbor based techniques
• Statistical techniques

Recently, reconstruction based techniques have emerged as a new kind of anomaly detection techniques.

DAGMM can be viewed as spanning both reconstruction based and clustering based techniques, which makes it more interesting. According to this paper, DAGMM comprises of two networks:

• Compression network: It consists of an encoder and a decoder. The encoder can reduce the input into a low-dimension representation and the decoder can reconstruct the input such that reconstruction errors can be computed. What's innovative is that the paper proposed to construct a combined representation consisting of both the low-dimension representation and two error features.
• Estimation network: Used to estimate the parameters for Gaussian mixtures based on the combined low-dimension representation. A sample energy for each low-dimension data point can be computed as an indicator for anomalies.

Dependencies

The code in this repo was implemented and tested with PyTorch 1.5.1 and Python 3.8.3