Clustering of radio emitter characteristics with complex-valued CNNs

Radio emitter identification has applications in the Internet of Things, spectrum monitoring and communication network security for the detection of emitters and for identification and authentication of transmission sources. Convolutional Neural Networks (CNNs) can learn the features of a signal from raw in-phase and quadrature (I/Q) samples for classification but can also extract the features of signals when used in inference. A CNN trained on signals from known emitters can extract the features of signals from unknown emitters. The extracted features when fed to a clustering algorithm can allow to group signals according to their emitters. This unsupervised learning technique can then identify signals from emitters unseen during training. Recently, models of complex-valued neural networks have demonstrated superior performance on data containing phase information in a variety of fields. This work aims to assess the ability of complex-valued neural networks to distinguish emitters unseen during training by comparing them to their real-valued counterparts. Our results indicate that complex-valued CNNs are superior to real-valued CNNs when trained with signals from emitters with similar extracted features.