With the rapid development of deepfake generative techniques, the synthetic hyper-realistic faces have attracted critical public attention. For the purpose of preventing privacy issues, many studies have been proposed to detect fake facial images. However, due to their poor generalization performance, existing deepfake detection approaches have encountered significant challenges posed by suspicious face generated by unknown synthetic algorithms. In this paper, we propose a novel identity disentanglement framework is devised to extract the overt and covert identities from the facial images and perform deepfake detection accordingly based on their consistencies. To formulate and monitor the disentanglement of overt and covert identities, we design reconstruction tasks that incorporate the face swapping tasks. Specifically, an identity encoder is designed to extract and disentangle overt and covert identities from the fake faces. Meanwhile, an attribute encoder is utilized to capture the facial attributes, such as age, gender, and expression. Then, a decoder is constructed to fuse attributes and identities and reconstruct the face-swapped images. The model ability in disentangling overt and covert facial identities is achieved by face swapping with different combinations of attributes, overt identities, and covert identities, restricted by a series of designed loss functions. Extensive experiments across multiple benchmark datasets demonstrate the superior detection performance with promising generalization capability of our approach compared to other state-of-the-art methods.