One-Class Classification for Anomaly Detection with Kernel Density Estimation and Genetic Programming

A novel approach is proposed for fast anomaly detection by one-class classification. Standard kernel density estimation is first used to obtain an estimate of the input probability density function, based on the one-class input data. This can be used for anomaly detection: query points are classed as anomalies if their density is below some threshold. The disadvantage is that kernel density estimation is lazy, that is the bulk of the computation is performed at query time. For large datasets it can be slow. Therefore it is proposed to approximate the density function using genetic programming symbolic regression, before imposing the threshold. The runtime of the resulting genetic programming trees does not depend on the size of the training data. The method is tested on datasets including in the domain of network security. Results show that the genetic programming approximation is generally very good, and hence classification accuracy approaches or equals that when using kernel density estimation to carry out one-class classification directly. Results are also generally superior to another standard approach, one-class support vector machines.