Federal Adversarial Domain Adaptation Model Based On FADA And Its Application In The Field Of Financial Credit

With the rapid development and wide application of big data and artificial intelligence technology,financial institutions represented by commercial banks continue to launch innovative credit products.However,in the early stage of production,innovative credit products are faced with the problem of cold start due to insufficient label samples and difficulty in training effective models.The traditional solution to this problem is to adopt the transfer learning method,that is,to use the abundant data of similar credit products of external institutions to assist in the construction of credit default prediction model.However,with the development of technology,cyber criminals constantly use new means to attack data,data privacy protection and security are facing more and more challenges.Banking industry,as a typical data-intensive industry,will be more restricted in the application of data modeling,resulting in the failure of data communication between different organizations,and the transfer learning method that requires direct access to external data is no longer applicable.Federated learning provides a new way to solve this problem.It can combine multiple data to build a model under the premise of protecting data privacy.However,when the data distribution difference is large,the model performance of federated learning will be inhibited due to the "negative transfer" phenomenon caused by data heterogeneity.Federated adversarial domain adaptation(FADA)can align the features of source domain and target domain through adversarial learning,alleviate the "negative transfer" phenomenon,and improve the model generalization performance on the premise of ensuring the accuracy of the model.However,the FADA algorithm needs to share the transformed feature among the devices.The feature has the risk of being decoded,which will bring significant risks to the banking industry,where data security protection is particularly important.Therefore,this paper proposes a privacy-protected FADA(P2FADA)model on the basis of FADA algorithm.On the one hand,this model replaces the shared feature in FADA with a shared gradient,which is more strict in data Privacy protection and has good applicability in the financial credit field that focuses on data privacy and security.On the other hand,the model alleviates the model accuracy decline caused by data heterogeneity through the counter domain adaptation technology.The model performance is better than the classical federal learning algorithm Fed Avg,and it is more efficient in solving the cold start problem of credit risk control.Through simulation experiments and empirical analysis,this paper shows that P2 FADA still has superior accuracy and stability under the scenario of data heterogeneity.Firstly,this paper designs data generation rules based on the credit default prediction process in real business scenarios.Secondly,two indexes,cosine similarity of conditional distribution and KL divergence of feature distribution,are selected to measure the difference between source domain and target domain from different dimensions,and it is proved that the accuracy of the model decreases with the increase of data distribution difference.Then,according to whether the feature distribution is the same and the condition distribution is similar between the source domain and the target domain,this paper sorts out four typical credit scenarios and conducts simulation experiments.P2 FADA has shown outstanding effectiveness and stability in the four experiments.Finally,this paper selects the credit data of Lending Club platform from 2015 to 2018 for empirical analysis.By comparing with logistic regression,XGBoost,ANN and Fed Avg models,the advantages of P2 FADA in solving the cold start problem of credit risk control and the phenomenon of "negative migration" are demonstrated.