Iris recognition is widely employed in unmanned detection environments, but it would suffer from a variety of attacks, such as, printed iris attack, wearing cosmetic contact lenses, glass eyes, prosthetic eyes, and so on. Although many previous methods have been proposed to resolve these problems, the generalization of models under cross-domain and cross-device scenarios is still need to be improved. To alleviate it, we propose an unsupervised domain adaptation transfer learning model with high detecting accuracy and generalization which is robust to attacks. The transfer learning is used to attain agile deployment in our model. Our model is mainly based on the CDD (Contrastive Domain Discrepancy) measurement method, which minimizes the intra-class difference and maximizes the inter-class difference, because of its satisfactory performance on liveness detection tasks. The transfer learning method based on MMD (Maximum Mean Discrepancy) measurement is proposed to attain agile deployment which selects the feature space alignment between the target domain and the source domain. Code is available at https://github.com/Wuxiuying111/Cross-device-iris-liveness-detection.git .