Automatic Modulation Classification using Shallow CNNs

Abstract

Automatic Modulation Classification (AMC) is the task of identifying the modulation scheme of an unknown communication signal, and it forms a critical component of cognitive radios, intelligent receivers, spectrum surveillance, and modern adaptive communication systems. AMC has been studied extensively, with early work focusing largely on Single-Input Single-Output (SISO) systems. Two broad families of techniques dominate the literature: the likelihood-based (LB) approach and the feature-based (FB) approach. The LB approach formulates AMC as a multiple-hypothesis testing problem, in which each of the candidate hypotheses corresponds to a distinct modulation type; while statistically optimal, it requires accurate prior information and is computationally demanding. FB methods are more efficient and need less prior knowledge, relying instead on discriminative features such as higher-order moments (HOMs) and higher-order cumulants (HOCs) extracted from the received signal.