Lightweight and Interpretable DL Model Using Convolutional RFF for AMC

The Automatic Modulation Classification (AMC) is important for effective communication in various scenarios. Traditional techniques face challenges with channel variability and complex deployment environments, requiring advanced solutions. Deep learning models have been applied to address these challenges, but they often lack interpretability, which is important for understanding and trusting their decision-making processes. This paper introduces the Convolutional RFF Threshold Denoiser Network (CRFFTD-Net), a novel deep learning model that not only enhances signal representation through an innovative application of convolutional Random Fourier Features (RFF) combined with a residual shrinkage building unit for noise reduction and a recurrent neural network for classification but also improves interpretability. This model significantly reduces the required parameters, decreasing computational load without sacrificing accuracy. Our approach provides a clear understanding of the classification decisions using Class Activation Maps analysis, demonstrating competitive performance against existing state-of-the-art models. We present comprehensive evaluations using the RadioML 2016.10A dataset, showing that CRFFTD-Net achieves high classification accuracy across various signal-to-noise ratios, making it a promising solution for next-generation communication systems.