Classifying Wi-Fi from Raw Power Measurements Using a Neural Network Adapted from WaveNet

Abstract

Research to identify coexisting wireless devices is becoming increasingly important, in part due to the yearly increase in internet users and wireless devices. Various machine learning algorithms, including neural networks, have been proposed and utilize a wide variety of data and feature extraction methods. Most leverage features from frequency domain because, although limited, these solutions tend to be less complex. In this thesis, a neural network that utilizes dilated convolutions is proposed to classify Wi-Fi standards using raw power measurements. The proposed model is adapted from a previous model, namely WaveNet [1], which was used for generating raw audio in text-to-speech (TTS) applications. With this method, synthesized audio sounds more natural than other state-of-the-art TTS methods. By utilizing dilated convolutions, WaveNet has a larger receptive field with few layers that can model long-range temporal dependencies. This serves as an advantage that both recurrent neural networks (RNN) and long short-term memory (LSTM) networks do not share. Wi-Fi power measurements are collected across 802.11n, 802.11ac, and 802.11ax wireless technologies both individually and with multiple technologies coexisting across a range of various throughputs. These are used to train the proposed model (WIFINet). Results indicate that 98.10% detection accuracy can be achieved by utilizing the proposed network. Investigations showed a convolutional neural network (CNN) with similar accuracy of 96.33%, indicating that modeling long-range temporal dependencies is not needed. At the very least, WIFINet yields little improvement over CNN for identifying wireless devices operating across the span of 802.11 technologies.