Deep learning has achieved remarkable results in modulation classification under two assumptions: a large amount of labeled class-balanced data is available, and test data shares the same distribution as data used for training. However, due to dynamic wireless fading channels and hardware imperfections, it is implausible that these assumptions hold in practice. This paper proposes Model-based Data Augmentation for Deep learning-based Modulation Classification (MDA-DMC) to build a high-quality dataset from a small amount of labeled seed data.
We develop two novel augmentation methods to combat channel and hardware impairments, along with two widely
used augmentation methods, adding Gaussian noise and rotation. We are the first to investigate the relationship
between augmentation methods and impairments due to channel and hardware imperfections. We show that MDA-
DMC can successfully combat the hardware imperfections for the same channel model as the channel model of
seed data. Furthermore, with a few AWGN seed data, MDA-DMC achieves an accuracy gain of up to 17.20% in fading channels. By adding a few fading seed data, MDA-DMC improves accuracy by an additional 10.78% in fading channels.