Unsupervised and nonparametric detection of information flows

The problem of detecting the presence of possibly bidirectional and time-varying information flows through two nodes in a network is considered. Only the transmission timing measurements are used in the detection. The proposed technique assumes no parametric flow model and requires no training data. The consistency of the detector is established for a class of non-homogeneous Poisson traffic. The proposed detector is tested in a simulation using LBL TCP traces (Paxson and Floyd, 1995 [24]) and an experiment involving MSN VoIP sessions.

► Detection of the presence of an information flow between two nodes in a network.
► We develop a nonparametric detector using only the transmission timings of nodes.
► The detector is memory efficient and has linear computational complexity.
► The consistency of the detector is proved under the Poisson traffic assumption.
► We test the detector using LBL TCP traces and MSN VoIP measurements.