Robustness of Deep Modulation Recognition under AWGN and Rician Fading

We study the robustness of modulation recognition using deep neural networks. This is of critical importance for applying deep learning for radio modulation classification, because wireless propagation conditions could vary significantly under different communication environments. We compare the performance of radio modulation recognition using data and using expert features. While deep modulation recognition using data has been proposed in existing literature since it achieves better performance than crafted expert features, our results indicate that using expert features yields significantly more robustness.