Research On Password Security Technology Based On Federated Learning

Password belongs to the privacy of users.It cannot be collected without telling the user.The relevant laws and regulations of the countries do not allow infringement of user privacy.Store a large number of user passwords on each server,in order to protect the user's password privacy,it is often stored for the hash value of the user password.However,establishing a password model requires the user password,and cannot use a hash value training.Thus,there is a contradiction between the privacy and deep learning training of the password.The purpose of this article is to solve the problem of difficulty in collecting password data.This article uses federated learning to train a password model,which can be trained to establish models without collecting user passwords,and protect users' privacy.This paper first designs a password model structure,namely,based on the password model of the Long Short-Term Memory network(LSTM),using the public password data set,through the traditional way,train a password model.Then,through the federated learning,the data is not directly collected,and the user's data is retained in the user,using two different federated learning architecture,training two previous structurally password models.Then,compare the model effects in the above training methods.According to the above,This article use the same data set and the same model structure,using deep learning and federated learning training models,respectively.Federated learning does not collect user passwords during the training process to protect users' privacy.At the same time,the password model of the fedrated learning training will be compared by the probability threshold graph,which is almost the same as the password model of traditional centralized training,and there is no significant difference.The above facts description: Through fedrated learning,it can establish the same password model with deep learning effects under conditions that guarantee user password privacy.The advanced nature of this paper is first manifested,in order to solve the difficult problem of collecting passwords,use a new method of fedrated learning;secondly,with semi-trust third part,simplify the secure aggregation process,can be directly used as a privacy safety protection module for federated learning;finally,through the probability threshold graph,compare the model effect of the above two training methods,There is no significant difference in the model effect of federated learning and deep learning training.