Inferring Origin Wi-Fi Traffic Patterns by Deep Learning of Max- pooled Sparse Representations of Received Frames

We introduce a novel networking application of deep learning to infer the origin traffic pattern in Wi-Fi from observing only received frame runs and gaps. CSMA adds linear and nonlinear altercations to the original characteristics of a flow, which is subject to further distortions from mixing with other flows. We devise a multi-layer inference pipeline based on sparse coding and dictionary learning popular in machine learning. In particular, we integrate max pooling at all layers to capture features (received frame characteristics) varying in length and position. We empirically evaluate the proposed pipeline in OPNET and present a comparative performance analysis with ARX/ARMAX least squares, Naı̈ve Bayes classifier, kernel density estimation, and EM-optimized Gaussian mixture model.