Learning the daily model of network traffic

Anomaly detection is based on profiles that represent normal behaviour of users, hosts or networks and detects attacks as significant deviations from these profiles. In the paper we propose a methodology based on the application of several data mining methods for the construction of the "normal" model of the ingoing traffic of a department-level network. The methodology returns a daily model of the network traffic as a result of four main steps: first, daily network connections are reconstructed from TCP/IP packet headers passing through the firewall and represented by means of feature vectors; second, network connections are grouped by applying a clustering method; third, clusters are described as sets of rules generated by a supervised inductive learning algorithm; fourth, rules are transformed into symbolic objects and similarities between symbolic objects are computed for each couple of days. The result is a longitudinal model of the similarity of network connections that can be used by a network administrator to identify deviations in network traffic patterns that may demand for his/her attention. The proposed methodology has been tested on log files of the firewall of our University Department.