Towards a Profiling View for Unsupervised Traffic Classification by Exploring the Statistic Features and Link Patterns

In this paper, we study the network traffic classification task. Different from existing supervised methods that rely heavily on the labeled statistic features in a long period (e.g., several hours or days), we adopt a novel view of unsupervised profiling to explore the flow features and link patterns in a short time window (e.g., several seconds), dealing with the zero-day traffic problem. Concretely, we formulate the traffic identification task as a graph co-clustering problem with topology and edge attributes, and proposed a novel Hybrid Flow Clustering (HFC) model. The model can potentially achieve high classification performance, since it comprehensively leverages the available information of both features and linkage. Moreover, the two information sources integrated in HFC can also be utilized to generate the profiling for each flow category, helping to reveal the deep knowledge and semantics of network traffic. The effectiveness of the model is verified in the extensive experiments on several real datasets of various scenarios, where HFC achieves impressive results and presents powerful application ability.