Mobile app traffic flow feature extraction and selection for improving classification robustness

Abstract

In machine-learning based mobile app traffic classification, flow feature distributions can easily drift due to changes in network environments, user habits etc. Unstable features may negatively influence mobile app traffic classification robustness, so to remedy this problem, this paper investigates how to obtain optimal feature sets for improving classification robustness of mobile app traffic. Specifically, we develop a method to search for stable and discriminative features by jointly analyzing mobile app traffic characteristics and assessing the degree of feature drift. Along these lines, we first analyze the in-flow behavior characteristics of traffic flows, so as to extract a potential feature set for mobile app traffic data. Next, we present two new metrics to assess the degree of drift experienced by the flow features from different perspectives and design a composite metric to score these features by considering the degree of drift as a penalty factor of discrimination power. Based on these metrics, we further propose an algorithm to search for optimal features with high discrimination power but low degree of drift. Existing flow features and feature selection algorithms were implemented for our comparison experiments. Our experimental results on real mobile app traffic data demonstrate the effectiveness of our feature set and feature selection algorithm on improving classification robustness.