Domain shift is a challenging problem in speaker verification, especially when dealing with unseen target domains. Recently, embedding decoupling-based methods have shown their effectiveness. Typically, domain information is extracted by a domain classification loss and then decoupled from speaker embeddings. However, the domain classification loss fails to ensure that only domain information is encoded in domain embeddings. This paper proposes a novel mutual information-based embedding decoupling framework, in which the domain information is extracted by maximizing the mutual information between different speaker sample pairs in the same domain. Then the domain information is removed from speaker embeddings by minimizing mutual information between speaker and domain embeddings. Experiments indicate that our method can improve the generalization and outperform domain classification-based decoupling methods.