Research On Data Aggregation Technology Based On Privacy-Preserving In Federated Learning

At present,Federated deep learning has been widely used in various fields,such as graphic image classification,autonomous driving,and speech recognition.However,in the process of training on large-scale data,there is a risk of leaking user sensitive data,and the computing efficiency of the operating equipment is insufficient to meet the training needs.Therefore,improving the capabilities of privacy protection and computing efficiency are two issues that need to be addressed in training.This thesis is based on the study of efficient and secure data aggregation methods that consider protecting the privacy of users in a federated learning environment.Through understanding and deep learning of existing data aggregation schemes based on Federated learning for privacy-preserving,the model framework of Federated learning,and various challenges faced by Federated learning are analyzed and summarized,including statistical heterogeneity,privacy protection,Communication overhead.Through research,it is found that the current privacy protection schemes under Federated learning rarely consider the fundamental problems of low data quality shared by some users(called low-quality data users)and different user equipment resources.Obviously,in the Federated training process,low-quality data may reduce the training efficiency and accuracy,and even cause the model to have no practical value.Therefore,when designing a Federated learning solution,in addition to considering user privacy,the heterogeneity of user equipment and user data should also be considered.This is a trade-off between privacy and training accuracy and efficiency.Aiming at the problems of current Federated learning,this paper considers using cryptography knowledge in a Federated deep learning environment to achieve the goal of protecting user data privacy.And proposes two solutions model to protect user privacy cloud-based Federated learning: SAHPP and SAHD.Among them,SAHPP considers the heterogeneity of user data and calculates the score of "data quality" for users to ensure that the global aggregate value is mainly based on some user data,which has a high contribution to training.As an enhancement scheme,the SAHD further optimizes SAHPP,improves the key agreement mechanism,we considers the heterogeneity of user equipment at the same time,and proposes the parameters of users' "reliability",which improves the accuracy and efficiency of model training.In addition,the safety and practicability of the two models are analyzed from the aspects of safety,training accuracy,and overhead.Compared with the existing schemes through experimental simulations,we find both the two schemes have higher accuracy and efficiency,and lower communication,storage,and computing overhead.In particular,the enhanced solution SAHPP not only considers the situation of heterogeneous data,but also considers users with lower equipment resources to participate in training,and uses a polynomial key negotiation mechanism to achieve a certain degree of efficiency while protecting user privacy.