Data sharing among financial institutions is often not possible, resulting in a "data silo" situation. In this paper, we train an efficient financial fraud detection model under the framework of privacy protection from the perspective of facilitating the collaboration of multiple financial institutions to train a fraud identification system. This paper first used traditional oversampling and under sampling methods to balance the data and train models such as logistic regression, support vector machines and random forests, but did not obtain the desired results. In contrast, the optimal subsampling method based on logistic regression performs well in terms of training results and program runtime. To protect data security, differential privacy is introduced on this basis to find the classification accuracy of the model under different privacy budgets. It is concluded that a certain balance between the degree of privacy protection and model effectiveness should be achieved according to privacy requirements.