Selection of On-line Features for Peer-to-Peer Network Traffic Classification

Abstract

The selection of features plays an important role in traffic detection and mitigation. Feature selection methods can significantly improve the computational performance of traffic classification. However, most of the selected features cannot be used online since these features can only be calculated after the completion of a flow. Another requirement is that an optimum number of features must be chosen so as to classify the P2P traffic within the minimumtime possible. Out of more than ten feature selection algorithms, it was discovered that Chi-squared, Fuzzy-rough and Consistency-based feature selection algorithms were the best for P2P feature selection. The proposed algorithm gives better feature subset for online Peer-to-Peer (P2P) detection using machine learning (ML) techniques. The process of validation and evaluation were done through experimentation on real network traces. The performance is measured in terms of its effectiveness and efficiency. The experimental results indicate that J48 classifier with online subset feature selection produces a higher accuracy (99.23