Outliers detection in network services with self-learned profiles

Wireless communication networks and services suffer from multiple kinds of security attacks which cannot be handled only at the wireless protocol level. This paper proposes an intrusion detection system that self-learns the user profiles using machine learning techniques. The system applies knowledge discovery techniques to generate a compact user profile offline. The profile is used to detect intrusions offline and online. Security breaches and ongoing attacks are identified detecting outlier activities in relation to the user profile and to immediate forecast behaviour. The later one provides a very fast warning, which is validated by the slower and more precise profile based online system. They are complemented by the slowest offline system, which is capable of maintaining updated user profiles. The system was implement using RStudio, and was tested using the 2014 Dendalion big data challenge dataset publicly available. The results show that the offline system has an outlier detection accuracy above 99% and that the online system was able to distinguish outlier activity from the users' own activity.