论文部分内容阅读
Recently,the security and privacy problems of artificial intelligence have attracted lots of attention from academia and industries.On the other hand,Federated learning is able to inhibit data’s leaking during the training process at certain level.Thus,Federated learning has a specific protective data security.Nevertheless,such security was not enough against strong adversary.There is still a chance of privacy leak in the federated learning model.To handle the privacy problem in federated learning,differential privacy has been introduced to make federated learning more secure with two types,which are local differential privacy with federated learning and a central differential privacy with federated learning.Both types have pros and cons.For federated learning with local differential privacy,all users enjoy with strong privacy and does not trust the server,the aggregator suffers from low utility.While for federated learning with centered differential privacy has high accuracy with low privacy.To overcome the above problem,we have studied the privacy in local differential privacy toward federated learning with shuffle.The overall framework consists of three stages,i.e.,users,shuffle,and analyzer.In this model,each user secures his local model by applying local differential privacy model and send the secure parameters as a message to shuffle,then shuffle receives a vector of messages from all users and then apply the random perturbations which change the messages index to hide the message position from server.Finally,the server aggregates the perturbed messages,which received it from shuffle.The practical results of shuffle is successful in solving the above problems.Our dissertation contains another new model from privacy,which named as a gradient boosting decision trees in shuffle model towards federated learning.In this model,we present privacy decision tree with shuffle model,where each user applies the local differential privacy during building his tree model then send the secure tree parameters to shuffle adding the randomly permutation before sending these messages to analyzer in order to aggregate it.In addition,the problem of unbalanced data has been studied by adding variable privacy budget for each user.For our best knowledge,our available studies deal with all datasets as balanced data.But,our proposed model applies the variable privacy budget ? for each user that depends on user dataset.Moreover,we demonstrated the results on both the real life dataset MNIST and the IRIS data,which demonstrate the improvement of our model for the privacy of last problems,which put the experimental foundation for the further study of the shuffle privacy in this active research zone.