Bayesian Neural Networks for Wi-Fi Device Fingerprinting and New Device Detection

David Justamante, Patrick McClure

Abstract: Recently, radio frequency (RF) fingerprinting using deep neural network machine learning models has gained popularity. However, these models are deterministic when making predictions. This leads to models that are notoriously poor at properly estimating prediction uncertainty. We propose using Bayesian neural networks as a method for improving the estimation of prediction uncertainty in Wi-Fi fingerprinting. In particular, we demonstrate that variational inference using spike-and-slab dropout results in neural networks that have improved test set accuracy, test set calibration, and accuracy for uncertainty-based detection of new devices not seen during training.