Machine Learning Based IP Network Traffic Classification Using Feature Significance Analysis

Abstract

After over 30-year deployment, IPv4 addresses are running short on supply with the growth of the Internet so dynamic. A new technology will take its place, IPv6, an evolution from IPv4 that includes virtually unlimited address space. However, it will take time to totally transit from IPv4 to IPv6. IPv6 will coexist with IPv4 for a period of time and then eventually replace IPv4. This paper studied network traffic that included information of both IPv4 and IPv6. The traffic was collected from 600 US government websites that were all reported to have Domain Name Services (DNS) and Web services accessible over IPv4 and IPv6. Cloud based, Internet distributed monitoring agents were deployed in eight geographic locations to collect data. Both feature selection algorithms, filter and wrapper, were applied to the dataset and the classification accuracy was then studied. The results showed that feature selection algorithms effectively reduced the complexity of the classification model. The results also confirmed that the reduced feature set contributed a superior classification performance over full feature set.

Keywords

• classification
• machine learning
• ipv4 and ipv6
• feature selection