Research On Social Network Differential Privacy Methods Based On Probability Models

With the development of the Internet,social network has become a field closely related to people’s daily life.The personal information is collected in large quantities,and the social network contains a large amount of personal information and sensitive data.As personal information is widely used,the privacy is facing the risk of disclosure.Nowadays,everyone has been faced by the threat of privacy disclosure,and the privacy of social network has also been widely concerned and studied.As a mathematically defined privacy protecting method,differential privacy is widely used in the privacy protection of social networks.However,as the different diversity of data types,social networks are not suitable for undifferentiated differential privacy operations directly.Therefore,constructing a probabilistic model to represent the network structure and performing differential privacy to the parameters of the probabilistic model are new ideas of differential privacy in social networks.Disturbing the parameters of probability models can prevent attackers from obtaining network models and network data,and realize the privacy protection of social network.Dynamic properties,edge weights and node attributes are not only the important properties of real social networks,but also the necessary research contents of differential privacy protecting method in social networks.Therefore,this paper takes different types of social networks as the research objects,combining probability theory and Bayesian theory,and studies the differential privacy protecting method of social networks based on probability models.The main research work of this paper includes the following four parts:Firstly,aiming at the problem of excessive noise in the differential privacy protecting method of evolving networks,and the problem of the invalid solution in the process of privacy calculation due to dynamics,a differential privacy protecting method of evolving networks based on Generalized Hierarchical Random Graph Model is proposed.In this method,the probability model of evolving network is constructed by using Generalized Hierarchical Random Graph Model,and the network structure of evolving network is abstracted.In order to reduce the amount of the noise,this method represents the dynamics model by locally adjusting the hierarchical structure by the Generalized Hierarchical Random Graph Model,and reduces the scale of the network structure which needs to add noise.At the same time,the method also calculates the connection probability of privacy nodes based on Bayesian theory to avoid invalid solutions in the process of privacy calculation.Experiments on real data sets show that the proposed method can satisfy the dynamic characteristics of evolving networks and the differential privacy,and improves the accuracy of data.Secondly,aiming at the problem of ignoring the privacy protection of the structure roles of the network nodes in the weighted network differential privacy protecting method and the excessive noise caused by the excessive edge weight,a weighted network differential privacy protecting method based on Random Block Model is proposed.Firstly,this method proposes a construction method of Random Block Model satisfying differential privacy,which can protect the privacy of the node structure roles.Then,according to this method,a privacy weighted network probability model is constructed.Based on the idea of Variable Bayesian,the privacy model parameters can be learned iteratively.In order to reduce the amount of noise added in the iterative process,an efficient calculation method of model parameter noise is proposed.Finally,experiments on real data sets show that the proposed method can satisfy the characteristics of weighted network,the differential privacy,and improve the accuracy of data.Thirdly,aiming at the problem of excessive noise caused by the association between topology and the node attributes in attribute network,an attribute network differential privacy protecting method based on probability model is proposed.Firstly,according to the idea of early fusion,this method combines the network topology and the attributes of nodes to construct the overall attribute network probability model.Then,the noise is added to the model parameters,and a synthetic attribute network satisfying differential privacy is generated through the privacy probability model.In order to further reduce the amount of noise added to the model,an improved privacy attribute network probability model is proposed.The model does not directly add noise to the model parameters,but constructs the hyperparameters of the model,and realizes the differential privacy by adding noise to the hyperparameters.As the scale of the hyperparameters is smaller,the amount of noise added to the model is also smaller.This makes the accuracy of the final output data better.Finally,experiments on real data sets show that the proposed method can satisfy the characteristics of attribute network,the differential privacy and improve the accuracy of data.Finally,aiming at the different effects of different attributes on the network structure in the differential privacy protecting method of multi-attribute networks,a differential privacy protecting method based on probability model is proposed.Firstly,for discrete multi-attribute networks,in the process of constructing the network probability model,the correlation parameters between multi-attributes and network structure are defined,and the factors that different attributes have different effects on network structure are added to the model.The metadata partition model is established by using correlation parameters.The combination of metadata is divided into independent groups,and the metadata is allocated through index mechanism.The differential privacy of discrete multi-attribute networks is realized by privacy model parameters and metadatas.Then,for continuous multi-attribute networks,Bernstein polynomial is used to approximate the probability function,and it obtains a stable expression of the probability function with slow changes with parameters.Finally,experiments on real data sets show that the proposed method can satisfy the characteristics of multi-attribute networks,the differential privacy and improve the accuracy of data.