Wireless network traffic disaggregation using Bayesian nonparametric techniques

We present a machine learning spectrum awareness framework capable of characterizing and inferring the application layer protocol states of multiple interleaved wireless network traffic flows using only externally observable energy detector features. The framework is intended to inform intelligent dynamic spectrum access (DSA) strategies in a cognitive radio environment. This extends an approach we developed previously for single isolated traffic flows, which applied a Bayesian non-parametric technique to construct hidden Markov model (HMM) representations of specific protocols. The learned HMM models, with hidden states closely corresponding to actual protocol states, were used for protocol classification and state recognition given a stream of energy detector observables from an isolated traffic flow. In this work, various single protocol HMMs are combined into a factorial hidden Markov model (FHMM) representing multiple heterogeneous interleaved flows. Using the FHMM to infer the states of the interleaved flows directly from observations of the aggregate traffic, we avoid having to deinterleave the transmissions of the component flows, a particularly difficult task in cognitive radio environments with agile emitters. We demonstrate this framework on an emulated network scenario with multiple simultaneous flows carrying different application layer traffic types.