Implementation of network traffic classifier using semi supervised machine learning approach

Network Traffic Classification using classical techniques such as port number based and payload based is becoming very difficult because many applications use dynamic port number and encryption technique to avoid detection. To overcome the drawbacks of classical techniques various machine learning techniques were proposed. Machine learning technique faces the problem of labeled instances (in supervised learning) and time consuming manual work (in unsupervised learning). To address the above problems we proposed a semi supervised machine learning technique. The key idea of proposed technique is to build a classifier from training dataset consisting of both labeled and unlabeled instances. For experimental purpose KDD CUP 99 intrusion detection dataset and MATLAB tool is used. We evaluate and compare the performance of the classifier build with 10%, 20% and 30% labeled instances in training dataset. The result of experiments show that classifier build with 30% labeled instances in training dataset has better performance at number of clusters equals to 50.