Clustering-Based Protocol Classification via Dimensionality Reduction

Abstract

We propose a unique framework that is based upon diffusion processes and other methodologies for finding meaningful geometric descriptions in high-dimensional datasets. We will show that the eigenfunctions of the generated underlying Markov matrices can be used to construct diffusion processes that generate efficient representations of complex geometric structures for high-dimensional data analysis. This is done by non-linear transformations that identify geometric patterns in these huge datasets that find the connections among them while projecting them onto low dimensional spaces. Our methods automatically classify and recognize network protocols. The main core of the proposed methodology is based upon training the system to extract heterogeneous features that automatically (unsupervised) classify network protocols. Then, the algorithms are capable to classify and recognize in real-time incoming network data. The algorithms are capable to cluster the data into manifolds that are embedded in low-dimensional space, analyzed and visualized. In addition, the methodology parameterized the data in the low-dimensional space.