Automatic Modulation Classification using Graph Convolutional Neural Networks for Time-frequency Representation

Recognition of the radio signal's modulating scheme is becoming increasingly important in civil and military applications. It can potentially alleviate the electromagnetic signal congestion in $5\mathrm{G}$ networks by utilization of dynamic spectrum access or perform friend/foe identification in electronic military warfare as well as to support the detection of cyber-security related attacks. The recent advances in graph-convolutional networks (GCN) reveal a potential for usage in applications such as automatic modulation classification (AMC). Considering the structure of the modulated signal in time and frequency, this work proposes GCN architecture for AMC in various signal to noise (SNR) levels. The experimental results reveal that such approach delivers comparable results to other approaches published in the literature.