Privacy-Preserving Research For Unsupervised Data Mining Methods

With the popularity of the Internet and mobile devices,the collection and storage of personal information has become increasingly easy.Data mining has become an important technology for extracting useful information and knowledge,providing valuable information and business opportunities for enterprises and governments.However,these data include a large amount of personal privacy information such as users’ personal identities,financial status,disease diagnoses,purchase records,social media activities,etc.As people increasingly use the Internet and mobile devices,privacy protection issues become more urgent.Therefore,effectively using this data while protecting privacy has become an urgent problem and important task.Clustering and manifold learning are two common unsupervised data mining methods that can automatically discover patterns,structures,and features from data,helping people better understand and utilize data.In various fields such as marketing,medicine,finance,etc.,these methods have been widely applied to data analysis and decision support,bringing many conveniences for business development and scientific research.However,these methods also face privacy protection challenges in use.Therefore,protecting user privacy and making these methods more secure,efficient,and usable are essential.This thesis aims to address privacy protection issues in clustering and manifold learning.The main research results are as follows:First,a privacy-preserving isometric feature mapping outsourcing scheme based on rotation transformation is proposed to address the privacy issues of non-linear manifold learning ISOMAP in outsourcing.This scheme fully considers the requirements of security,efficiency,and data availability.The main innovations of this scheme are: firstly,a data encryption scheme for multi-user vertical partition data sharing mining is designed,and users can customize the degree of privacy protection.Secondly,a multi-user ISOMAP mining scheme under a semi-honest environment is proposed,and all outsourcing mining processes are completed by the cloud server without user online participation.Thirdly,through performance analysis in multiple aspects such as trustworthiness,continuity,geodesic distance error,downstream task accuracy,and visualization,the utility of this scheme is verified.Experimental results show that this scheme can calculate on ciphertext data,and the effect is almost the same as that of the original data.Secondly,a privacy-preserving clustering similarity evaluation scheme is proposed to address the privacy protection issues in clustering similarity calculation.The main innovations of this scheme are as follows: firstly,a clustering similarity privacy protection scheme based on the co-association matrix or consensus matrix is designed,which can be applied to clustering and clustering integration.Secondly,in order to protect the privacy of the co-association matrix or consensus matrix,we adopt geometric transformation and differential privacy protection methods respectively.Thirdly,we compare the clustering similarity errors of different privacy protection methods through 19 common clustering similarity evaluation indicators as utility evaluation,analyze the impact of different privacy protection methods on clustering similarity indicators and the sensitivity of clustering similarity indicators to interference.Through the research in this thesis,we have explored the privacy protection issues in unsupervised data mining methods and proposed innovative solutions.These solutions aim to provide more secure,efficient,and usable data mining methods,and provide valuable references for future research.