Unsupervised Emitter Clustering through Deep Manifold Learning

Abstract

We perform unsupervised clustering to group Radio Frequency signals according to the device that transmitted the signal. We do so by first performing supervised training on a RiftNet classifier that contains a layer that can be used as a latent vector. We then perform the unsupervised clustering on a completely different set of devices than those used for training. During unsupervised clustering, we project each signal into its latent vector, then perform clustering in a UMAP-learned reduction of that space. This approach provides understanding and a path forward on how such semi-supervised deep-learning clustering approaches might fit in a real world RF system.