Automatic modulation classification facilitates many important signal processing applications. Recently, deep learning models have been adopted in modulation recognition, which outperform traditional machine learning techniques based on hand-crafted features. However, automatic modulation classification is still challenging due to the following reasons. Existing deep learning methods are only applicable to the data of the same distribution. In practical scenarios, data distribution is varying with sampling frequency, thus domains with different sampling rates are formed. Besides, it is difficult to construct large-scale well-annotated datasets for all domains of interest. We define the domain with sufficient data as the source domain, while the domain with insufficient data as the target domain. Obviously, the classification model performs weakly in the target domain. To address these challenges, we propose an adversarial transfer learning architecture (ATLA), incorporating adversarial training and knowledge transfer in a unified way. Adversarial training performs an asymmetric mapping between domains and reduces the domain shift. Knowledge transfer is used to mine prior knowledge from the source domain. Experimental results demonstrate that the proposed ATLA substantially boosts the performance of the target model, which outperforms the existing parameter-transfer approach. With half of the training data reduced, the target model achieves competitive recognition accuracy to supervised learning. With one-tenth of training data, the promoted accuracy is up to 17.3% points.