Cross-domain fake news detection, aiming to detect fake news in unseen domains, has achieved promising results with the help of pre-trained language models. Existing approaches mainly relied on extracting domain-independent representations or modeling domain discrepancies to achieve domain adaptation. However, we found that the relationship between entities in a piece of news and its corresponding label (fake or real) fluctuates among different domains. Such discrepancy is ignored by existing methods, leading to model entity bias. Therefore, in this paper, we propose a novel cross-domain fake news detection method based on dual-granularity adversarial training from the perspective of document-level and entity-level. Specifically, both the news pieces and their entities are modeled individually to construct an encoder that can generate domain-independent representations using adversarial training. Moreover, the dual-granularity soft prompt, consisting of two independent learnable segments trained on the source domains, is employed to make the model easily adapt to the unseen target domains. In addition, MultiFC, a released dataset for cross domain fake news detection, is not suitable for the evaluation due to its unreasonable domain construction rules. We artificially reconstructed the dataset and named it New-MultiFC, which is a more domain-discriminative dataset. Experimental results on both the newly constructed New-MultiFC and FND3 show the effectiveness of the proposed approach, achieving the state-of-the-art results in unseen domains.