Research For Privacy-preserving Domain Adaptation

How to circulate and share data,and give full play to its factor value to be used in social production and life in more places is an important issue in the digital economy era.In an era where data privacy is more and more concerned,how to make a large number of data resources on the Internet be used for machine learning and deep learning models without revealing privacy can effectively promote the value of data to be fully exploited.Domain adaptation research has made great progress in recent years on solving the problem of how to improve the performance on target domain of the machine learning model trained on source domain,where the source domain and target domain encounter the distribution inconsistency.However,existing domain adaptation methods require the source domain data and target domain data to be co-stored in the same place and at the same time during training,which is likely to be unsatisfactory in practical applications.Because of data privacy constraints,source domain data and target domain data often must be strictly separated and invisible to each other.This thesis adopts the differential privacy framework to analyze the training of domain adaptation,and uses methods such as density estimation and distribution approximation to propose a framework that enables domain adaptation algorithms to ensure data privacy during training.This thesis mainly does the following two tasks(1)In order to solve the privacy problem of source domain data,this thesis adopts the method of approximating the distribution of source domain features,and proposes a domain adaptation framework that can strictly guarantee the privacy of source domain data.The process of this framework supports model training under the condition that the source domain and target domain data are completely separated,and there is no coordination requirement for the two parties.The method proposed in this thesis is perpendicular to the existing domain adaptation research.All domain adaptation methods can use our framework to ensure the privacy of the source domain data during the training process.In addition,this thesis also designs a domain adaptation algorithm.The algorithm makes the target domain model training process without using any source domain related information other than the source domain model.The framework proposed in this thesis is experimentally verified on four different datasets,and the results show that our method can obtain accuracy close to that of existing methods on domain adaptation tasks under the premise of protecting the privacy of source domain data,but with a privacy guarantee will bring a certain degree of accuracy decline,and when the amount of data and calculation is sufficient,the loss of accuracy caused by the privacy guarantee can be ignored.(2)In order to solve the privacy problem of target domain data,this thesis proposes a domain adaptation framework that can strictly guarantee the privacy of target domain data by using the method of distribution modeling of target domain features.During the training process,the source data party,the target data party and the computing party can be completely separated,and the computing party can complete the training of the model without directly touching the target domain data.This process strictly guarantees data privacy.This can serve as a solution to the problem of how data on the Internet containing private information can be effectively used to train models.In this thesis,experiments are conducted on three different datasets to verify the effectiveness of the proposed method,and the results show that with sufficient computation or data volume,our method can perform domain adaptation tasks while ensuring data privacy.The results are close to those of existing methods.