This paper presents a novel approach for unsupervised domain adaptation (UDA) that utilizes task-specific classifiers to align feature distributions between source and target domains. The proposed method aims to address two key issues with existing UDA methods: the lack of consideration for task-specific decision boundaries and the difficulty in matching feature distributions due to domain characteristics. To achieve this, the method maximizes the discrepancy between the outputs of two classifiers, which are trained to detect target samples far from the source distribution. A feature generator learns to generate target features near the source distribution to minimize this discrepancy. The effectiveness of the method is demonstrated through experiments on image classification and semantic segmentation datasets, showing superior performance compared to other UDA methods. The code for the method is available at <https://github.com/mil-tokyo/MCD_DA>.This paper presents a novel approach for unsupervised domain adaptation (UDA) that utilizes task-specific classifiers to align feature distributions between source and target domains. The proposed method aims to address two key issues with existing UDA methods: the lack of consideration for task-specific decision boundaries and the difficulty in matching feature distributions due to domain characteristics. To achieve this, the method maximizes the discrepancy between the outputs of two classifiers, which are trained to detect target samples far from the source distribution. A feature generator learns to generate target features near the source distribution to minimize this discrepancy. The effectiveness of the method is demonstrated through experiments on image classification and semantic segmentation datasets, showing superior performance compared to other UDA methods. The code for the method is available at <https://github.com/mil-tokyo/MCD_DA>.