Analysing Variational Autoencoder-based Anomaly Detection with Biometric Data

The broad field of anomaly detection is relevant for almost every aspect of life. We as humans are able to identify anomalies because they might pose a threat or be crucial for survival. Automated anomaly detection is used in any kind of production, in autonomous systems and many more areas. Unsupervised anomaly detection plays a huge role because anomalies come in many different shapes and for some domains it is almost impossible to know how anomalies will look like. Real world data is unlabelled and because it often is high dimensional too this work’s approach utilizes a variational autoencoder to gain a representation with lower dimensions on which the anomaly detection is performed. This works analyzes the combination of the variational autoencoder and common anomaly detection methods. This work was able to hint at a transfer of this method which has been successfully applied to other domains to biometrics, but is not able to establish a reliable way of applying the idea. A very challenging problem and the crux of this work is the concept of the anomaly itself as not even humans
from the same cultural background agree with each other on this topic.