Research On Privacy Protection Algorithm Based On Generation Model

Generation models are widely used for data generation and data enhancement in the field of data scarcity because they can approximate the joint probability distribution of training data sets.Differential privacy has become one of the most effective privacy protection methods because of its strict theoretical proof,which provides an effective technical means to prevent privacy leakage in data mining.Thus,it has become a hot topic in generative model research.However,the existing privacy strategy based on generative model has several problems:different noise design and injection location have a great impact on the effectiveness of the model,the consumption of privacy budget is large and the training time is long.This paper carries out systematic studies aiming at the balance between model utility and security,and the consumption of privacy budget.The main contents include how to understand the potential impact of noise design on the model,how to determine a reasonable gradient clipping threshold,and how to make the generated data contain more attribute features.The main work and achievements of this paper are as follows:(Ⅰ)Propose an adaptive differential privacy protection algorithm based on GAN.The algorithm used GAN to generate the training data set,add gaussian noise to the gradient through differential privacy mechanism in the learning process of GAN,and update the gradient threshold in an adaptive manner.Meanwhile,this paper introduces moment account to automate the tracking calculation of privacy loss.Compared with the conventional exponential function update strategy,the new strategy can update the threshold according to the current gradient value and public training sample iteration gradient.Therefore,the size of the injected noise can be flexibly controlled,which enhancing the availability of the resulting data.Test results not only proves the relationship between privacy protection degree and output image quality of generator,but also proves that the algorithm can achieve the purpose of privacy protection while ensuring data availability.(Ⅱ)Design a distributed InfoGAN privacy protection model based on auxiliary network.This model starts with the training process of model generation and adds a generator on the basis of traditional InfoGAN to solve the problem of pattern collapse.By learning potential representation,it can better capture the attributes of training data set.Then this paper makes a distributed implementation of InfoGAN differential privacy protection model.For each client,the privacy protection mechanism is trained and introduced,and the global sensitivity and noise mechanism are determined.Thus,the various attributes of images owned by different clients can be captured in a way of privacy protection.The model reduces the number of parameter exchanges by using auxiliary network.It only need to transfer parameter between client discriminator and the auxiliary neural network,which greatly reduces the communication load and is more cost effective.Test results on MNIST datasets show that the model can generate images with more thickness and Angle variation than the nondistributed environment.