A New Framework for Automatic Modulation Classification using Deep Belief Networks
Pejman Ghasemzadeh, Subharthi Banerjee, Michael Hempel, Hamid Sharif · 2020
Automatic Modulation Classification (AMC) is the process of determining the modulation scheme of an intercepted signal with no a priori information about its characteristics. AMC's main advantage is that no communication overhead needs to be allocated for control information to inform the receiver about changes in a transmitted signal's modulation scheme. Proposed approaches for AMC traditionally suffer from inherent computational complexity that prevents real-time AMC applications. Several extensions for lowering the computational complexity have been developed, including Feature-based AMC with machine learning classifiers. For the contribution of this research, we propose a new classifier called Deep Belief Network (DBN) for AMC applications, which is an algorithm derived from the Restricted Boltzmann Machine (RBM). DBN is capable of learning the probability distribution over its set of inputs and is comprised of layers of RBM stacked together. Additionally, this composition of layers in DBN leads to a faster learning procedure compared to conventional algorithms, and can help this classifier operate in real-time. Essentially, High-Order Statistics-based (HoS) features are utilized in the feature extraction stage for which the authors also investigate the bias issue of the estimator. The standard RadioML dataset is utilized to assess the performance of the studied AMC platform. This research shows the comparison of high order modulation schemes results in lower-bound performance of the AMC classifier, which are classified with notable higher probability of correct classification compared to conventional deep learning classifiers especially for lower signal-to-noise (SNR) ratio scenarios.