As the advance of multimedia technology, the sources of information are not limited to one form but emerging multi-modal properties. Through the multimedia data processing, it is an essential component how to detect the fake news to avoid their spread. The fake news can leverage multimedia contents to fabricate evidences or mislead readers, which damages a lot for data management in social networks. In this work, we explore the task of multimodal fake news detection. The major challenge of fake news detection stems from the modality fusion by abundant information. Overcoming the limitations of the current models, we tackle the challenge of learning corrections between modalities in news, and substantially propose a block-based representation and fusion mechanism that can leverage the middle knowledge both from the spatial and temporal respective. Except for the original fake news detector, our model outperforms others from two components: pooling blocks-based feature extractor, attention blocks-based feature aggregator. The performance of our proposed architecture is evaluated on Weibo dataset, which indicates the superiority over the baseline model to an evident degree.