Research On Key Technologies Of Privacy Protection On Social Network

In recent years,Online Social Networks(OSN)connect a large number of users.Users could communicate with friends,show their profiles and share text and pictures with others on social networks.Social network service providers collect data generated by users and provide users with personalized services.However,social network data often contain sensitive information,because they are closely related to users' personal lives.Unfortunately,there are untrusted social network service providers which sell users' sensitive data to third parties for surplus profit.What's worse,even if the social network providers add noise to perturb the sensitive data before releasing the data,the attackers can still use the de-anonymization methods to attack the identity information of some users or use the inference attack methods to predict the sensitive information.Therefore,it is an urgent problem to hide the sensitive information generated by the complex behaviors of social network users.There have been some works about social network privacy protection.But the exist-ing works still have shortcomings.Generally,there are three disadvantages of the existing methods:First,the interaction among users is not considered.Second,the global information is not utilized to launch privacy attack well.Third,there is no protection mechanism against inference attack on social network embeddings.To overcome these disadvantages,the dissertation provides three research works as follows:(1)Risk analysis of users' sensitive information leakage.This dissertation proposes a three-party game framework that includes users,a social network service provider,and an adversary to model the social network private data trading.At first,to observe network evolution after user's privacy is leaked,the complex interactions of users are mod-eled.Then this dissertation integrates the network evolution results to the three-party game framework.After that the data trading between the service provider and adversary is formulated to be a Nash bargaining game,for which Nash bargaining solutions are analyzed via both theoretical analysis and numerical experiments.The analysis of this dissertation can clearly illustrate data trading strategies between the service provider and adversary and offer guidance for designing privacy protection schemes.(2)Seed-free social network de-anonymization.To solve the problem of not fully uti-lizing the global information to launch privacy attack,this dissertation proposes a seed-free social network de-anonymization method.This method is different from the traditional seed-free deanonymization method which uses local and manual features.In our method,a deep neural network is adopted to learn features automatically and an adversarial frame-work is employed for node matching based on global features.To be specific,the latent representation of each node is obtained by graph autoencoder.Furthermore,an adversarial learning model is proposed to transform the embedding of the anonymized graph to the latent space of auxiliary graph embedding,such that a linear mapping can be derived from a global perspective.Finally,the most similar node pairs in the latent space as the anchor nodes are utilized to launch propagation to de-anonymize all the remaining nodes.The extensive experiments on some real datasets demonstrate that our method is comparative with the seed-based approaches and outperforms the start-of-the-art seed-free method significantly.(3)Privacy-preserving social network embedding against inference attack.In order to overcome the lack of protection mechanism for middle term of deep learning-based social network analysis,this dissertation proposes two social network embedding methods against inference attack.Firstly,this dissertation proposes privacy-disentangled and privacy-purged mechanisms based on adversative learning,both of which can independently remove sensitive attribute information from the latent representation.Then,we combine these two mechanisms to an end-to-end model and propose a graph embedding model for property privacy protection.Finally,following the idea of the privacy-disentangled mechanism,this dissertation also proposes a link privacy-preserving graph embedding model.Extensive experimental results on real-world datasets show that the proposed models are superior to the start-of-the-art methods in both privacy protection and utility preservation.