Research On Enterprise Credit Evaluation Model Based On Blockchain And Federated Learning

Accurate enterprise credit evaluation can create a new pathway for the information gap between investors and operators and enhance market dynamics.To address the problem of data privacy leakage brought by centralized modeling in existing enterprise credit evaluation models,this paper proposes an enterprise credit evaluation model based on blockchain and federated learning,which unites multiple enterprises to perform model training without data out of the local area through distributed training of federated learning and ensures the privacy security of data.To address the problems of malicious training and model performance degradation of clients in distributed training,an incentive mechanism based on the credit mortgage value is designed in conjunction with application scenarios to motivate enterprises to actively participate in federated learning,perform reliable local training,and achieve quality control of enterprise credit evaluation models.The specific research work in this paper is as follows.(1)To address the privacy leakage problem in modeling the enterprise credit evaluation model,this paper proposes an enterprise credit evaluation model based on blockchain and federated learning.Firstly,the distributed training of federated learning is used instead of the centralized training method in traditional machine learning to ensure the privacy security of enterprise data.Secondly,through blockchain technology,the process of model distribution,aggregation,and incentive of federated learning,as well as the model is consensus and secure.Finally,the modeling is completed based on two different types of machine learning algorithms,Fed-ANN and Fed-XGB.Through comparative analysis of simulation experiments,the Fed-ANN-based model has a significant advantage in accuracy when the sample size of data owned by the enterprise is large enough,and the model based on Fed-XGB has the advantage of faster convergence;when the sample size of data owned by the enterprise is too small,the model based on Fed-XGB outperforms the model based on Fed-ANN in terms of accuracy and convergence speed.(2)For the malicious training behavior of clients in the model modeling process of blockchain and federated learning,an incentive mechanism based on the credit mortgage value is designed in conjunction with application scenarios for motivating clients to actively participate in federated learning training and perform reliable training locally.Firstly,the client is assigned a credit mortgage value based on its historical credit rating,then the quality of the client’s local model is verified,and the client is penalized or rewarded accordingly by the quality verification results.The simulation results show that the incentive mechanism has an incentive for the client to train reliably locally and provide high-quality local models;when the client’s malicious training probability increases to 30%,the clients’ required credit mortgage value to participate in the federated learning increases to 6.2 times and the gain decreases by 114.2%.(3)In order to realize the quality control of enterprise credit evaluation models,the model quality control methods based on the credit mortgage value are applied in the modeling process based on Fed-ANN algorithm and Fed-XGB algorithm,respectively.By setting a certain malicious training probability for enterprises and comparing and analyzing the changes of accuracy and convergence speed of enterprise credit evaluation models,the results show that the model quality control method based on the credit mortgage value can achieve quality control of enterprise credit evaluation models and can reduce the impact of malicious training of enterprises on enterprise credit evaluation models.It can improve the accuracy and reduce the convergence time of the models.