Differential Privacy Protection Method Based On Federated Learning

With the rapid development of artificial intelligence and the popularity of mobile devices,the application scenarios that require the cooperation of multiple participants are constantly emerging,and the role of distributed data processing and distributed ma-chine learning is increasingly prominent.For example,financial data scattered in mul-tiple banks,medical records in different hospitals,behavior records of each user under the large platform,and data generated by smart meters,sensors or mobile devices need to be processed and mined in a distributed manner.Data island is one of the important challenges faced by distributed data processing and distributed machine learning.As a so-lution to data island,federated learning is a promising distributed computing framework,which can train models locally on multiple decentralized edge devices without transfer-ring their data to the server.With the improvement of citizens’ awareness of privacy and the improvement of relevant laws,privacy security issues in federated learning are also increasingly concerned by people.And the latest research work shows that it has been able to attack the gradient parameters of the model and restore the user’s privacy data,which means that it is not enough to protect privacy only by keeping the data local.In addition,privacy protection technology protects privacy at the expense of model accuracy.For this reason,this paper uses differential privacy technology to protect the privacy of users in federated learning,and analyzes the convergence nature of the model for distributed sce-narios,selects more reasonable parameters,so as to reduce interference noise and achieve the purpose of improving the accuracy of the model.The main work of this paper includes the following aspects:· Transaction level differential privacy model based on federated learning In order to solve the problem that the existing privacy algorithms usually need to im-prove the model privacy at the expense of model accuracy,thus reducing the model availability,we optimize the model from three different perspectives under the dis-tributed scenarios: First,we propose a more compact upper sensitivity boundary for the distributed scenarios.Second,we propose a better one than the traditional average allocation of privacy budget.Thirdly,the weight is allocated according to the amount of noise to reduce the overall noise impact.· A hybrid differential privacy model based on federated learning Based on the above algorithm,a hybrid differential privacy model is further proposed to allocate the privacy model on-demand and reduce the noise impact of the global model.In addition,we also analyzed the convergence of the federated learning algorithm un-der the differential privacy mechanism and proposed an improved method according to the two error terms of training,namely the tailoring value learning method and the improved combination method.· Federal learning system based on differential privacy protection To solve the privacy security problem in model training,we introduced the improved differential privacy technology into the federated learning framework to implement the feder-ated learning system of privacy protection,which is used to protect the data privacy of the client during the training process.Firstly,the parameter server generates the initial global model and distributes it to each client.Secondly,after the local train-ing of each client,the corresponding differential privacy noise is added according to the client’s privacy needs.Finally,the parameter server aggregates the client’s model,and if there are users who trust the server,the differential privacy noise is injected to protect their privacy.