Prediction Of Privacy-sensitive Regions Based On Features Of Movement And Network Of Mobility Data

With the development of GIS and radio-communication network technology,LBS applications are developing rapidly.Within these applications,we can get users’ mobility data with tremendous economic and social benefits.However,when combining these data with sensitive semantics,mobility data will be privacy-sensitive.Attackers can obtain users’ privacy information via analyzing sensitive mobility knowledge.Thus,it is a major research to protect user privacy while ensuring the availability of mobility data.The most crucial point while designing corresponding protection method is to find sensitive regions.Traditional methods are classification via spatial data topology computing and feature classification based on remote sensing image.However,it is difficult to obtain map data with sensitive semantic.Meanwhile,feature classification methods are easily affected by the resolution of remote sensing images.Therefore,we proposed a supervise classification method based on Spark big data platform aiming at recognize privacy-sensitive regions via analyzing movement features and network features of sensitive regions.The specific steps mainly include:(1)Matched user mobility data with spatial grid to obtain movement features of each grid.(2)Built a network with spatial grid unit and topology each grid with its neighbor grids to obtain network features.(3)Established classification models.Topology spatial grids with specific geographic features with sensitive attribute and got classification tags of each grid.Combined classification tags with movement features and network features,made supervise classification based on Spark MLlib machine learning platform.(4)Predict sensitive attribute of each grid based on classification model.Using training models to classify test data and calculate prediction accuracyIt is proved:(1)the accuracy,f-measure vale and AUC of a prediction become greater with the increase of training data and test data ratio.However,all evaluation indicators of the classification showed a downward trend when the ratio changed from 8:2 to 9:1.(2)The prediction results obtained via our methods considering multi-features are more accurate comparing with methods that consider only one.