Résumé
Data mining for intrusion detection can be divided into several subtopics, among which unsupervised clustering has controversial properties. Unsupervised clustering for intrusion detection aims to i) group behaviors together depending on their similarity and ii) detect groups containing only one (or very few) behaviour. Such isolated behaviours are then considered as deviating from a model of normality and are therefore considered as malicious. Obviously, all atypical behaviours are not attacks or intrusion attempts. Hence, this is the limits of unsupervised clustering for intrusion detection. In this paper, we consider to add a new feature to such isolated behaviours before they can be considered as malicious. This feature is based on their possible repetition from one information system to another. We propose a new outlier mining principle and validate it through a set of experiments.