Abstract: Interpretable fake news detection aims to identify a piece of news as true or false with a reasonable explanation. Existing interpretable fake news detection models usually adopt sequential models to model texts, and use the information in the original reports to assist the models in judging whether the news is true or false and generating explanations. In contrast, this paper proposes an interpretable fake news detection model of dual-encoder evidence distillation neural network, which uses Gumbel-Softmax to replace previous TOP-K technique, so that the model can adaptively select valuable reports. We also introduce a graph neural network and a dual-encoder structure. The experimental results obtained on two public datasets verify the effectiveness of the proposed model.