A Robust Graph Convolutional Neural Network-Based Classifier for Automatic Modulation Recognition

The procedure of automatically recognizing the modulation scheme of the received signal without any knowledge of the communications parameters employed by the transmitter has gained tremendous attention for developing various applications such as dynamic spectrum access in next generation communications systems and in electronic warfare (EW) applications. Amongst the proposed approaches for this procedure, the feature-based approach working on the principles of deep learning models has consistently demonstrated more favorable properties for its integration with real-world applications and realtime operations. However, this approach faces the challenges of low performance efficiency caused by low classification accuracy and high computational complexity in environments characterized by low signal-to-noise (SNR) ratio conditions. Therefore, in this paper, we present our research findings on a deep learning-based classifier for Automatic Modulation Recognition (AMR) with robust classification accuracy and lower computational complexity, thus leading to higher overall performance efficiency. The feature extraction stage of the proposed classifier operates on the translation of received signal constellations into a graph, which then follows a process of extracting graph information. Mapping the extracted information into features is implemented by a convolutional block. This design employed in our feature extraction stage is the main element for the robustness of our proposed classifier. For our performance evaluations we selected 8QAM, 64QAM and 256QAM, representing candidates for low-, medium- and high-order modulation schemes, respectively. Our simulation results exhibit higher classification accuracy compared to related literature works, by an average of 11 percentage points (p.p.), 11.31 p.p. and 9.3 p.p., respectively, for the successful classification of 8QAM, 64QAM and 256QAM across an SNR range from −10 to 30 dB. The computational complexity analysis indicates that our proposed classifier is capable of executing the AMR task with lower processing latency. Its robust classification and lower execution latency combine into a more efficient overall performance provided by our proposed classifier.