A Robust Perceiver-Based Automatic Modulation Classification for the Next-Generation of Wireless Communication Networks

Automatic modulation classification (AMC) is an indispensable part of intelligent receivers in modern wireless communication systems. AMC enables blind identification of modulation without prior knowledge of the signal parameters, which is a challenging task, particularly in practical scenarios with severe multipath fading, frequency-selective and time-varying channels. Although deep learning techniques have been shown to be efficient in AMC tasks, traditional convolutional and recurrent neural networks may not be able to cope with complex-valued input signals and large-scale datasets. Motivated by this, in this paper, we propose a novel perceiver-based AMC architecture that leverages the recently introduced Perceiver, which combines cross-attention and latent transformer modules, to efficiently process and classify complex-valued in-phase and quadrature (IQ) samples of the received signal. The proposed model is trained and evaluated on the DeepSig 2018 RadioML dataset. Simulation results demonstrate a significant improvement in the classification accuracy compared to a ResNet-based AMC model, particularly for higher-order quadrature amplitude modulation (QAM) and under practical signal-to-noise ratio values. These findings indicate the potential of the perceiver architecture for robust and efficient AMC in wireless communication systems.