Abstract: Although existing face anti-spoofing methods perform well in intra-domain testing, their performance significantly degrades in cross-domain scenarios. Current cross-domain face anti-spoofing methods based on domain adversarial alignment cannot guarantee that the alignment task directly serves the classification task since the alignment and classification networks are independent of each other. We propose a domain adaptation for face anti-spoofing method based on domain invariant liveness features dual disentanglement and progressive adversarial alignment. Firstly, the source domain features are heuristically disentangled into domain specific features and domain invariant features. Then, the gradient of classifier is used to perform a second disentanglement of the live-related and live-unrelated features in the domain invariant features. To alleviate optimization difficulties during training, a curriculum learning method is adopted to progressively align target domain features and the combination of live-related and live-unrelated features, gradually increasing the proportion of live-related features, and enhancing the correlation between the target domain features and face anti-spoofing task. From a causal perspective, we provide an explanation for the liveness alignment domain adaptation. Experimental results on CASIA-MFSD, Idiap Replay-Attack, MSU-MFSD, and OULU-NPU datasets demonstrate that the proposed method achieves the best average HTER value of 22.5% compared with ten existing methods and the current state-of-the-art performance on four evaluation protocols. Especially the HTER values of I-M and O-M evaluation protocols achieve 12.4% and 12.8%, respectively. The proposed method can significantly reduce the error rates of the model in the target domain and has better cross-domain generalization ability.