Research On Credit Risk Prediction Based On Federated Learning

In recent years,with the rapid rise of global Internet finance,financial technology with artificial intelligence technology as the core has deeply influenced people’s lifestyles,and the development mode of traditional financial industry has also undergone tremendous changes.At present,the field of credit risk is faced with the problems of scattered user data sources and unbalanced data distribution.On the one hand,financial institutions have few data sources and poor quality,so it is difficult to train high-quality models;on the other hand,there will be a risk of sensitive information leakage when training data.People also gradually pay attention to personal privacy and data security,and the state has higher requirements for laws and regulations.Although Internet finance has made people’s live faster,it also poses a huge challenge to credit risk modeling capabilities.Based on deep learning model related to the field of credit risk assessment,this paper combines federal learning technology to solve the problem of centralized computing power distribution,and constructs a federal deep learning model by using multi-source and multi-end data for risk prediction.The details of this research are as follows:(1)By combining ResNet(Residual Network,ResNet)with federated Learning technology,a.Unlike traditional credit risk modeling methods,the FL-ResNet can achieve multi-customer collaborative training and improve the performance of the prediction model while aggregate credit risk prediction model based on ResNet and federated learning.Firstly,the SMOTE+ENN hybrid sampling method is used to balance the loan data set.Then,the training of ResNet is placed on each client using the horizontal federal learning algorithm framework,and the central server is responsible for aggregating these local model parameters.Compared with the model training of single data,the FL-ResNet model has better prediction effect and higher safety.(2)To address the communication cost problem in the federal training process,based on the FL-ResNet model,this study uses the Local Global federated Averaging algorithm(LG-Fed Avg)is used to optimize the federated training process,and a new credit risk model(LG-FL-ResNet)is constructed.The model can reduce the number of communication global parameters without additional computational complexity and has lower communication costs,which can improve the efficiency of financial institutions.(3)Finally,this study uses the loan default public data set of Baidu’s paddle platform to verify the performance of the credit risk prediction model constructed in this paper,and sets up a large number of comparative experiments.The experimental results confirm that the federal credit risk prediction model has good feasibility and superiority in practical applications.Federal learning combined with artificial intelligence to prediction the credit risk of loan users.On the one hand,it can use financial data for rapid prediction and accurate prediction to help financial institutions improve operational efficiency and business processing speed,On the other hand,it can realize financial intelligence and make new explorations and attempts for the research and development of smart finance.