One-Against-All Methodology for Features Selection and Classification of Internet Applications

Abstract

Traffic classification by Internet applications, even on off-line mode, can be interesting for many applications such as attack identification, QoS prioritization, network capacity planning and also computer forensic tools. Into the classification problem context is well-known the fact that a higher number of discriminators not necessarily will increase the discrimination power. This work investigates a methodology for features selection and Internet traffic classification in which the problem to classify one among M classes is split in M one-against-all binary classification problems, with each binary problem adopting eventually a set of different discriminators. Different combinations of discriminators selection methods, classification methods and decision algorithms could be embedded into the methodology. To investigate the performance of this methodology we have used the Naïve Bayes classifier to select the set of discriminators and for classification. The proposed method intends to reduce the total number of different discriminators used into the classification problem. The methodology was tested for classification of traffic flows and the experimental results showed that we can reduce significantly the number of discriminators per class sustaining the same accuracy level.