Nonparametric Anomaly Prediction and Discovery

Abstract:

In this talk I'll discuss two variations on the problem of using training data to perform anomaly detection. In anomaly prediction, the data are all realizations of a "typical" or "normal" distribution, and the task is to design a rule for deciding whether future observations are anomalies. In anomaly discovery, the data may be a mixture of typical and anomalous observations, and the task is to identify the anomalies in the given data. I will discuss a general nonparametric framework for addressing these problems, while offering control on the false-alarm rate, based on the idea of minimum volume sets. Theory as well as practical algorithms based on support vector machines will be presented, and an application to network anomaly detection will be given.