Research On Privacy-preserving Methods Against Link-prediction-based Attacks

Link prediction analyzes the topology of social networks and predicts the potential relationship between two individuals.Deep learning improves the performance of link prediction.Even if the data publisher removes the user's sensitive relationship in the published data set,the attacker can still use the link prediction algorithm to predict the user's sensitive relationship in the social network,resulting in the user's privacy leakage.The link prediction method based on deep learning brings new challenges to the research of privacy protection methods.Therefore,the data publisher needs to protect the privacy of the published data,and the defense link prediction technology predicts the sensitive relationship,while ensuring the utility of the published data.In view of the privacy problem caused by using deep learning method to predict links,the current related research draws lessons from the anti attack method of deep learning,and uses the method of generating anti attack samples to cheat the link prediction method of deep learning,which leads to wrong prediction of sensitive links and protects the privacy of sensitive links.However,the proposed methods are not suitable for large-scale social networks,lack of judgment basis of disturbance effect,and do not consider the actual situation of protecting multiple sensitive links at the same time.In order to solve the shortcomings of the existing methods,this paper proposes a local disturbance privacy protection method,and proposes a privacy protection method for multiple sensitive links.The main work and contributions of this paper are as follows:(1)In view of the situation that the existing research is not suitable for large-scale social networks,LDIG algorithm reduces the disturbance range by dividing the closed subgraph of sensitive links,only calculates the integral gradient of the links in the disturbance range,reduces the time complexity and is suitable for large-scale social networks.(2)In view of the lack of judgment basis of disturbance effect in the existing research,the LDIG algorithm does not take the number of link disturbances set in advance as the judgment basis of ending disturbance,but takes the wrong prediction result of the link prediction method on the sensitive link in the disturbance graph as the judgment basis of ending disturbance,which reduces the number of link disturbances required for sensitive link protection,and improves the utility of privacy preserving publishing graph.(3)In order to protect multiple sensitive links,this paper proposes a privacy protection method based on CUDA parallel framework on the basis of local disturbance algorithm LDIG,which is divided into host side and device side.The host side uses CPU to complete the disturbance range division,link disturbance sorting and link disturbance stage sequence execution,The device uses GPU to complete the parallel processing and acceleration of each stage,and uses less time to generate the release graph to protect the privacy of multiple sensitive links in large-scale social networks.(4)The experimental results of four real network data sets show that the proposed algorithm is less complex than the previous method,which is suitable for privacy protection of large-scale social networks.It can reduce the number of link disturbances while providing link privacy protection,and ensure the effectiveness of data,and the algorithm is universal.The algorithm in this paper uses CUDA parallel framework to greatly reduce the time of generating the release diagram of multiple sensitive links.