The outbreak of COVID-19 has brought huge challenges to health systems around the world. Computed Tomography (CT), as an effective method to screen COVID-19, can provide explainable visual information which can be used for the diagnosis and quantification of COVID-19. It will make diagnosis easier and more effective by providing segmentation results of infections from given CT scans. Deep learning based methods can perform well in tasks that have abundant annotated data, while they may suffer from great performance drops when there is little annotated data available. To collect and annotate CT scans of COVID-19 cases is time-consuming and labor-intensive, which will demand highly-proficient radiology expertise. To tackle this bottleneck, this paper provides a novel unsupervised domain adaptation method that bridges the gap of distributions between the available annotated CT scans and private CT scans, thus reducing dependence on private annotated CT scans. Quantitative analysis demonstrates the effectiveness of our proposed method, which achieves ～ 20 increments in Dice score compared to models without adaptation.