Research And Implementation Of Differential Privacy Protection Technology Under Federated Learning

Differential privacy technology protects numerical data and non-numerical data through a noise mechanism and an exponentmechanism.It uses strict mathematical proofs and quantitative analysis to effectively protect private data.Traditionally,differential privacy protection is performed on centralized data.But now facing the situation of distributed data,privacy protection needs to be carried out in the environment of federated learning.Federated learning aims to establish a federated learning model of distributed data sets.At the beginning of its design,it faced the problem that cross-domain,cross-department,and cross-industry sensitive data sets could not be centralized.Thus,it needs to process locally and then build a learning model collaboratively.This thesis focuses on the privacy issues in the multi-party collaborative construction model in the federated learning environment,and mainly does the following three aspects:(1)In the face of the data analysis using XGBoost(e Xtreme Gradient Boosting)under federated learning,to address the privacy leakage problem that may occur during the construction of weak classifiers,a differential privacy protection algorithm based on XGBoost is proposed.This algorithm first splits the nodes in the weak classifier with the best segmentation point,using an exponential mechanism for differential privacy protection.Then it perturbs the leaf nodes through the Laplacian mechanism to protect the classification results.Therefore,in the process of constructing the model,the privacy and the security protection can be both carried out while the effect of data efficiency can be also achieved.The experimental results show that under the differential privacy protection based on XGBoost,compared with the traditional random forest differential privacy protection,it has better accuracy under the premise of taking into account the privacy protection.(2)Aiming at the need for multiple participants to jointly construct a model that provides privacy protection in a federated learning environment,a multi-party collaborative construction solution based on federated learning to meet differential privacy is proposed,which solves the problem of deploying XGBoost algorithm within a federated learning environment in the case of multiple participants.Also,the privacy protection was carried out during the deployment process.In the case of multi-party participants having only their own data sets,the model uses the aggregation server to pass its own split point to negotiate with other participants,decide the most suitable split point,and finally build a model collaboratively.Experimental results show that the model has the characteristics of short running time and high efficiency under the premise of ensuring the accuracy.(3)This thesis designed and implemented a prototype system of differential privacy protection technology under federated learning.The differential privacy algorithm module are implemented in the federated learning environment.,With the medical data set of a hospital,experiments are conducted using the prototype system.The experimental results showed that the prototype system can not only ensure data security,but also have high classification accuracy.rate.