Radio Frequency Fingerprinting Using Autoencoder Generated Features on IEEE 802.15.4 Networks

This paper studies the efficiency of Radio-frequency fingerprinting (RFF) on IEEE 802.15.4 networks using properties extracted by machine learning techniques. As a test case, we used the measurements of 33 wireless sensor devices to create a dataset with 640 complex samples from the signals' synchronization header of the frames transmitted by all sensors in two moving scenarios. This paper evaluates using an autoencoder (AE) neural network (NN) for extracting the RFF features. It addresses the challenge of identifying the performance of different AE NN architectures using varying numbers of features obtained from the AE latent state (LS). We compare the relative performance of the AE NN with varying types of NN (combining dense, convolutional, and recurrent layers) with five different LS dimensions (between 4 and 128), as well as two classifiers: a multi-layer perceptron (MLP) and a random forest (RF). An optimization for the encoder's LS, which uses the weights of the AE and a classifier NNs on an additional NN training phase, is also proposed. We show that a larger LS size does not always lead to better classification accuracy and that the AE loss is a bad predictor for the classifier performance. The best F1-score achieved was 93%, measured using the MLP classifier and the optimized one-dimensional convolutional AE with an LS dimension of 8, or 86% using the same configuration without the optimization. The RF classifier is much faster to train and still achieves an 88% F1-score for the optimized Gated Recurrent Unit (GRU) AE with an LS dimension of 16.