Analysis of Open Set Deep Neural Network Variants towards Classification of Known and Unknown Signals

In environments crowded with electromagnetic activity, cognitive radios (CR) must have the ability to differentiate between different signals, so as to understand their origins if needed, and decide whether they are friendly or unfriendly signals. This is extremely important in the case of military and intelligence applications, but is beginning to become important in the commercial world as well, given the recent discussion around 5G and ORAN. In this paper, we develop and compare three open and three closed-set signal classification models based on various Machine Learning (ML) architectures. Results show that open-set networks are able to perform very close to their closed-set counterparts while also being able to classify unknown signals that they have not been trained on.