Research On Privacy Preserving And Spam Filtering Technology For Extensive-Social Networks

With the widespread popularity of smart mobile terminals,the rapid developments of various social networking applications providing users with convenient and fast experiences of communication and sharing,while the increasing number of users apparently make the social network become much more extensive,which has a bad effect on the sharing and exploring of the users.When users share their own status,the growing friends-list with more and more unfamiliar people leads to the wider spread of personal information without intention,resulting in more risks of privacy leak.Furthermore,due to the mix of good information and bad information on the internet,users are often disturbed by spam including advertisements while browsing information,which does great harm to the normal experience of the user.For the above issues,this thesis proposes a privacy control framework based on the division of communities with the quantification of intimacy in the sending end and an upgradable spam filtering system in the receiving end.Specifically speaking,the main work and contributions of this thesis are as follows:1.A homogeneity walk based scheme of fuzzy community detection is proposed.The core idea behind the scheme is to learn the vector representation of the node in the network and then get the division of communities according to the clustering of the similarity between nodes.By dividing the communities of the friends-list,when the user is going to publish his status information,the status information is selectively pushed to friends in the relevant group,therefor achieving the goal of controlling the spread of the private information.Specifically,the concept of the homogeneity walk is proposed at first,which makes the process of the random walk tend to discover the homogeneity in the network structure and let nodes with tight connections in the social network get closer in the vector space.Then,after using the fuzzy clustering algorithm,the membership degree of each group is computed by the membership degree matrix and the threshold is set,making a node be able to belong to multiple communities at the same time,which solves the shortcomings of traditional schemes that cannot be used for overlapping community detection.Finally,the evaluation in the real-world social network proves that this scheme makes a significant improvement to the internal modularity and the external accuracy of the discovered communities.2.An intimacy-based privacy permission setting scheme is proposed.In order to overcome the shortcomings of the binary definition of friend relationship in the social network that cannot accurately measure the intimacy between friends and has poor flexibility,by using the data mining techniques,this scheme provides a fine-grain method to quantify the intimacy according to the attribute information,behavioral data and environmental factors of the friend and it can set the corresponding access level based on the quantified relationship to further control the range of the private information spread.Specifically,the influencing factors of the intimacy between friends is analyzed at first and the quantified relationship model is trained in the offline state.Then the intimacy is computed in the online state using the real-time characteristic data acquired in the online social networking platform.Finally,different open levels of privacy permissions are set according to the computed relationship value.The experiment shows that the scheme is effective and is able to be applied practically.3.An upgradable spam filtering system is designed.In the beginning,the embedding matrix of words is generated by the Word2 vec model using the text data obtained from the social network.Then the designed LSTM-CNN algorithm is used to improve the classification by combining the property of processing the timing information and the capacity of feature extraction in long short-term memory and convolutional neural network.The upgrading system is designed to constantly make the classifier be adjusted to the changes in text characteristics by re-marking fuzzy samples.Compared to traditional classification algorithms through experiments,the filtering scheme proposed in this thesis can achieve better results in accuracy.