The proliferation of multi-modal fake news poses an increasing threat to society. To deal with this issue, several algorithms have proposed to detect fake news by infusing multimodal information. However, the high-dimensional nature of multimodal features constrains the efficiency of fake news detection. To address this challenge, this paper proposes a Multimodal-Feature-masked Networks for fake news detection. Specifically, a learnable dimensional mask is proposed to adaptively reduce the interference of irrelevant information, effectively alleviating the adverse effects of high-dimensional features. Additionally, to utilize text-image consistency as a crucial criterion for classification, we devise a two-stage training strategy for cross-modal consistency detection and news veracity classification to enhance detection accuracy. Through extensive experiments on real-world datasets, the proposed approach is validated to achieve higher accuracy and less runtime compared with the previous work.