Prof. Josef Kittler presents Anomaly detection: A novel framework

Anomaly is an important notion in the operation of both biological and
engineering systems. The concept refers to events or situations which deviate
from normality (usual observation, order, form or rule) and in this sense are
considered anomalous. Applications include intrusion detection in surveillance
systems and communication networks, abnormality detection in medical
diagnostics, novel class detection in pattern recognition, and out-of-vocabulary
word detection in speech recognition, to mention just a few examples.

Anomaly is a very generic term, having many nuances, which are impossible to
discern using simply the classical mathematical concept of anomaly as an outlier
of a probability distribution. A recently proposed anomaly detection system
architecture is presented and discussed. It includes several distinct mechanisms
to detect anomalous events and facilitates their characterisation. In addition
to the conventional process of distribution-outlier detection, the mechanisms
include classifier incongruence detection, data quality assessment, classifier
confidence gauging, and model-drift detection. The outputs of these processes
feed into a reasoning engine, which draws conclusions about the presence of
anomaly and its nature. The advocated approach to anomaly detection is
illustrated on a number of applications.