The rapid spread of low-cost rumors in social networks has brought significant negative effects on society. At present, the detection models for fake news often introduce entity information in knowledge graph for enhancement, ignoring the problem that the actual detection accuracy decreases when the model relies on the training set entity for prediction. To solve this problem, a new Entity Deficiency Fake News Detection framework EDFND(Entity Deficiency Fake News Detection) is proposed. Firstly, entity knowledge is obtained from the article based on external knowledge and zero-sample entity linking technology. Secondly, a multi-level GRU based on the content of the article is designed to enhance the interactivity within and between sentences, and a dual attention mechanism is proposed to weaken irrelevant entities, and combine original information with entity information representation for fake information detection. Finally, the EDFND model is verified, the experimental results show that the EDFND model has better detection accuracy than the similar fake information detection models, and it is verified that the problem of decreasing detection accuracy can be solved well by reducing entity dependence.