点击次数：

发表时间：2021-01-01

发表刊物：Jisuanji Xuebao/Chinese Journal of Computers

所属单位：电子信息与电气工程学部

卷号：44

期号：10

页面范围：2090-2103

ISSN号：0254-4164

摘要：In recent years, with the vigorous development of technology and its related industries, Internet finance has increasingly highlighted its advantages. For a long time, qualification review based on the user information has been a fairly important business in the financial field. In most cases, when an individual applies for a loan from a bank, the bank will evaluate him or her through the actual situation based on the established predictive model to determine whether to grant the loan. In this process, a high-quality default risk assessment can avoid unnecessary losses for the banks. However, there are still many deficiencies in the current research on the assessment of default risks of borrowers by banks and other lending institutions. On the one hand, it is difficult to build a high-quality prediction model due to the lack of user data; on the other hand, people are paying more and more attention to the privacy protection of personal data, it is also tough work for banks to obtain a large amount of relative data, and because of that, they cannot carry out the prediction models to accurately predict users' situation. In order to solve the problem of joint modeling in the case of data is not shared, this paper introduces the idea of thefederated learning to effectively utilize the value of other participants' data without the leaving of local data to establish a shared predictive model. Because decision tree algorithms are widely used in financial risk controlling and fraud identification, this paper proposes a decision tree algorithm FL-DT (Federated Learning-Decision Tree) based on federated learning. Federated learning is the concept put forward by Google in 2016, which can complete joint modeling without data sharing. Specifically, the data of each owner will not leave the local place, and the global sharing model will be jointly established through the parameter exchange method under the encryption mechanism in the federal system (in the case of not violating data privacy protection regulations).Moreover, each participant only serves for the local targets. Firstly, a data storage structure based on a histogram is presented for communication transmission, which can effectively improve training efficiency by reducing the number of communications. Secondly, the garbled Bloom filter based on an oblivious transfer is proposed to perform the privacy set intersection, and then we can obtain the federated histogram containing the data information of each participant, and establishes the federated decision tree model. Finally, amulti-party collaboration prediction algorithm is put forward to improve the prediction efficiency of FL-DT. Based on four commonly used data sets in the financial field, this article assesses the accuracy and effectiveness of the FL-DT algorithm. The experimental results show that the prediction accuracy of the FL-DT model is higher than that of the model established using only local data, which is close to the model built in the case of data concentration. In addition, the prediction accuracy of the FL-DT methods is better than other federated learning methods, and the training efficiency and prediction efficiency are also better than other algorithms. © 2021, Science Press. All right reserved.

备注：新增回溯数据