Generative Adversarial Network-Based Signal Inpainting for Automatic Modulation Classification

Automatic modulation classification (AMC) aims to automatically identify the type of a detected signal in an intelligent wireless receiver, such as software-defined radio (SDR). Recently, deep learning-based methods convolutional neural networks have been applied AMC, showing high-accuracy performance. However, earlier studies do not consider various degradations that can possibly occur during transmission and reception signals. Particularly, be often unstable, partially received due dynamic spectrum sensing or systems. The corrupted with missing samples considerably degrades accuracy models, because it is very different from training datasets. To address this issue, preprocessing process restoring signal, called inpainting, essential. Although significant for classification, no performed investigate effect inpainting on AMC. end, study proposes generative adversarial network(GAN)-based method fills signal. proposed restore time-domain up 50% while maintaining global structure each type. correct recovery enables extraction distinctive features play key role classification. we perform intensive experiments RadioML dataset has widely used AMC studies. We compare performance two state-of-the-art models without respectively. Through analysis results, show GAN-based significantly improves