Optimization And Application Of PATE Model Based On Federated Learnin

The Private Aggregation of Teacher Ensembles(PATE)scheme is an important framework for privacy protection in the field of machine learning.However,it has the problem of low classification accuracy of teacher models caused by independent data storage and data division.Federated learning is a machine learning framework for protecting data privacy.Federated learning jointly trains a federated model on local data of every party,which can be used to solve the problem of low model classification accuracy in PATE.Therefore,this paper utilizes the idea of federated learning to optimize the PATE scheme for improving the classification performance of the model.However,there is still the problem of model parameter leakage by federated learning training the teacher model.To address this problem,this thesis proposed the privacy protection scheme of the PATE model by combining federated learning,differential privacy,and homomorphic encryption.Finally,this dissertation designed and implemented a credit card fraud detection system.The specific research contents are as follows.(1)This thesis proposed an optimization scheme of PATE model based on horizontal federated learning.In PATE,the robustness of the model is guaranteed by training hundreds of teacher models.However,the independently stored data leads to a low classification accuracy of the model during training.To this end,the idea of federated learning is introduced to train the teacher model,the federated average algorithm is used to improve the classification accuracy of the model,and the Gaussian mechanism is used to enhance the privacy protection of the model.Theoretical analysis results show that the proposed optimization method can improve the classification performance of the model.(2)This thesis proposed a privacy protection scheme of PATE model based on horizontal federated learning.It brings about the risk of privacy leakage by using federated learning to train teacher models.To solve this problem,introduced homomorphic encryption and differential privacy technology in the process of federated model training,and removed the differential privacy mechanism in the prediction aggregation stage.The experimental results show that this proposed scheme can improve the classification accuracy of the model and protect data privacy.Moreover,the proposed scheme can reduce the influence of the number of teacher models on the training results.(3)This thesis designed and implemented a credit card fraud detection system.Since the data is independently stored in different institutions in credit card fraud detection,it is a typical distributed machine learning application scenario.This paper designed and implemented a credit card fraud detection system based on mainstream development technologies such as Tensor Flow and Django.The system mainly includes four functional modules: information management,data management,model training,and model using.The system test results show that the proposed system can identify credit card fraud with a high accuracy rate,and can achieve the expected goal.