Autism Spectrum Disorder (ASD), as a developmental disorder of brain, affects the ability of individuals to express themselves verbally, participate in social activities and perform normal behaviors. Multi-site dataset inevitably introduces experimental and environmental variability in data acquisition and processing, which is not disease-related. For the purpose of reducing the impact of site effects and utilizing the connection between different functional community of the brain, a harmonization method is used to process the feature matrix and the brain topology metric nodal local efficiency is introduced as a weighting coefficient for feature enhancement in this paper, based on which a transformer architecture is developed to incorporate a community-interaction module. The result shows that our method achieves an accuracy of 73.4%, an AUROC of 79.97%, a sensitivity of 68.1% and a specificity of 78.63% in the binary classification task of ASD identification on the Autism Brain Imaging Data Exchange (ABIDE) dataset.