Lightweight Hybrid Deep Learning Scheme for Automatic Modulation Classification in Cognitive Radio Networks

Abstract

Addressing the spectrum scarcity problem under the explosive growth of wireless network applications and data traffic, requires a decidedly global research focus on Cognitive Radio Networks (CRNs), which have been widely regarded as the front-runner to cope with spectrum inefficiency. In this paper, the emphasis is given to Automatic Modulation Classification (AMC) as a means to blindly identify the underlying modulation format of intercepted signals in Cognitive Radio Networks. Modulation recognition is a necessary task to accomplish a set of Cognitive Radio functions, such as dynamic spectrum access and interference detection. Recently, several Deep Learning (DL) schemes have been explored to address the existing limitations of conventional approaches in dealing with the AMC problem. DL techniques have delivered outstanding performance in terms of reliability, scalability, and applicability compared to traditional methods. However, when it comes to resource-constrained end-devices, most of these techniques are not supported due to their high processing and storage requirements. In this work, we propose a hybrid DL model with attention mechanism based on layers of bidirectional gated recurrent unit (BiGRU) and convolutional neural network (CNN) architectures. Our model has a memory footprint of 126.6 kByte, slightly smaller than the memory footprint of the baseline model with 100 GRU cells (128 kByte), and achieves good performance on different scales of training data.