<p>Automatic modulation classification (AMC) aims to recognize the modulation type of a received signal, playing a crucial role in various civil and military applications. However, most existing deep learning (DL)-based AMC methods require massive labeled samples to ensure usability, which limits their application in many practical scenarios. Although some few-shot learning (FSL) methods have emerged to address the limited data problem, they often struggle to generalize to new classes due to differences in feature distribution. In this paper, we propose a novel AMC method that simultaneously addresses the challenges of limited data and domain distribution differences. Firstly, we introduce an advanced signal transformation method for mapping time-series signals to easily classifiable images. Secondly, we design an efficient convolutional neural network (CNN) for feature extraction, incorporating several feature-wise transformation layers to align features across different domains. Then, we utilize a few-shot graph neural network (GNN) with a well-designed optimization strategy to construct a large number of training tasks, endowing the model with robust anti-domain shift capabilities for few-shot classification. Numerical experiments are conducted on the RadioML2018.01A and RadioML2016.10A datasets. The simulation results validate the effectiveness of the proposed input mixing approach and demonstrate that our cross-domain few-shot AMC method outperforms existing few-shot methods.</p>