Analyzing User Privacy Preference And Risk Based On User Behaviors On Social Media

Social media refers to the Internet platforms on which users are allowed to share their generated contents or to communicate with friends,which are nowadays closely related to people lives,entertainment,and work.Some platform systems provide the access control mechanism by privacy settings to help users manage their personal information and published data.The platforms providers often set privacy settings to "public" for users as default so as to attract more audience.Since most users are less sensitive about the privacies on these platforms,the information they release is open available under such settings.By taking advantage of these data,attackers can analyze user identities,attributes,locations and other sensitive personal information,which brings a huge risk to user privacy.This dissertation tries to help user better manage their behaviors on social media so as to reduce the privacy risk.The contributions are summarized as follows:(1)To satisfy user requirements on fine-grained privacy management,this dissertation proposes the personalized recommendation method on user privacy policy that models the user privacy preferences as the probabilistic associations between items and friend links.The existing methods on privacy management consider much on the user similarities about their attributes,friend links,and the behaviors of open contents,and less on the fine-grained authorizations on the features of subjects and objects.We propose the privacy quantification method that models the complex relationships between the authorized subjects and objects as the probabilistic associations.Since the tags on objects are defined by users and may be inconsistent,we adopt the Folksonomy to learn the semantic characteristics of the objects.Based on the learned user preference,each privacy setting on a specific object is computed against the attributes of authorization subjects and objects,which satisfies the user expectation on fine-grained privacy requirement.Experiment results on real datasets show that the proposed approach recommends user privacy policies more accurately and more efficiently compared with the existing methods,which is especially important for online usages.Besides,it is applicable for the authorization to a new visitor.(2)We propose the user identity linkage method based on the publicly available behaviors,which reveals the potential risks on user identity privacy.The publicly available behaviors are the information that user released,such as the comments and attitudes on items,which can be accessed by all internet users.The existing methods mainly analyze the rich information,such as user attributes,relationship networks,and user published contents,by either the statistics method or the graph methods.However,it is difficult to obtain these informative user data due to the restriction of privacy settings.Different from these methods,we make use of the publicly available data to infer user sensitive information.Facing the characteristics of fragmentation,random noise,and low information content,we introduce the semantic knowledge base to analyze the implicit semantics of user behaviors.We also adopt the object contents that are associated with user behaviors to enrich user information.Then user identities are modeled from different aspects,including the explicit features of behaviors,the context of behaviors,the associated item information,and the implicit semantics of behaviors.We also combine these multiple aspects for identity linkage.Two real datasets on social media comments are used for experiments.The results show that the proposed method is more effective in identifying user identities.These results can help user understand their privacy risk in publicly available behaviors.(3)We propose attribute inference method based on the cross-platform user behaviors so as to analyze the privacy risk on user attributes on social media platforms.The existing studies mainly utilize the available user attributes and friend relationships to deduce user attributes on a single platform.But in fact,even the low risk on one platform does not mean the same low-risky on multiple platforms,which is often overlooked by users.This dissertation proposes the cross-platform user attribute inference method by embedding the implicit features of users and the consistency of anchor users cross platforms.We introduce the Ontology knowledge to model the hierarchical semantics of concepts and adopt the labelled data on the auxiliary platforms to create the identity mapping function between platform users,by which we can use more auxiliary information for user attribute inference.For the case with few cross-platform users,we propose another meta-learning method for the cross-platform attribute inference.Experimental results on Sina New and Sina Weibo datasets verify the effectiveness of our methods.(4)Aiming at the location privacy issue of social media users,this dissertation analyzes the location sensitivity problem based on the public check-in data.The existing methods mainly utilize the behavior information or the scenario data on check-in to predict the sensitive label of location by feature engineering,where this informative information are not always available in real applications.By only using the simple check-in sequences on locations and a very few location labels,we predict the semantic labels on the unmarked locations so as to evaluate the sensitivity of user location.The predication model embeds the locations and labels as the vectors in the same latent space,with the help of the context in both the check-in sequence and the label sequence,as well as the consistency constraints on location and its associated label.Based on the embeddings of locations and labels,the label can be predicated using the distance measurement of embedding vectors.Two datasets on the typical check-in platforms are adopted for comparative experiments.The results show that,compared with the baselines,the proposed method achieves higher accuracies on label prediction,which can provide users the more knowledge on location sensitivity.