In this paper, we propose a ResNet based neural architecture to solve the problem of Automatic Modulation Classification. We showed that our architecture outperforms the state-of-the-art (SOTA) architectures. We further propose to binarize the network to deploy it in the Edge network where the devices are resource-constrained i.e. have limited memory and computing power. Instead of simple binarization, rotated binarization is applied to the network which helps to close the significant performance gap between the real and the binarized network. Because of the immense representation capability or the real network, its rotated binarized version achieves $85.33\%$ accuracy compared to $95.76\%$ accuracy of the proposed real network with $2.33$ and $16$ times lesser computing power than two of the SOTA architectures, MCNet and RMLResNet respectively, and approximately $16$ times less memory than both. The performance can be improved further to $87.74\%$ by taking an ensemble of four such rotated binarized networks.