Research And Application Of Privacy Preserving Based On Discriminant Component Analysis

With the continuous development of artificial intelligence and big data technology,a series of problems such as data abuse and user privacy leakage emerge endlessly,which has attracted widespread attention.In some application scenarios with micro devices(such as intelligent wearable devices),in addition to the privacy problem,due to the limited resources of micro devices,their data processing ability is relatively weak.The emergences of compressive privacy and compressed sensing technology provide a new way to solve these problems.In addition,data may be lost or polluted by noise in the process of transmission and storage,which will seriously affect the utility of data.Matrix completion method can complete incomplete data or restore contaminated data,which is widely used in recommendation system,image in painting and other fields.This thesis focuses on the problems in the above application scenarios,and the main work includes:(1)A compressive privacy scheme based on matrix completion and discriminant component analysis is proposed.In this scheme,in order to avoid the leakage of the original information,the data is compressed(desensitization),and then uploaded to the public server.Desensitized data can only guarantee the performance of utility tasks,and cannot be applied to privacy tasks.In addition,in order to improve the performance of utility task,the missing original data is supplemented.Specifically,in the optimization objective of matrix completion,the discriminant information regularization term of maximizing utility and minimizing privacy is introduced.Through alternating iteration,the projection matrix of maximizing utility and minimizing privacy information is obtained while completing the data.Finally,the feature subspace is used to compress and desensitize the data.The experimental results show that the proposed scheme can achieve the expected goal.(2)A compressive analysis scheme based on Discriminant component analysis is proposed.Compressive sensing has great attraction in wireless sensor networks,which can extend the service life of resource constrained devices,it can also provide certain privacy protection ability.Existing analysis schemes based on compressive sensing have the problems of high reconstruction complexity,low accuracy of utility task classification,and privacy security.Therefore,the proposed scheme aims to avoid the highly complex reconstruction process,explore the possibility of analysis in compressed domain,at the same time,the compressed privacy projection dimensionality reduction technology is introduced to further protect user privacy.The proposed scheme includes two stages: off-line stage and online stage.In the off-line stage,learn projection matrix,compute discriminant transformation matrix of compressed privacy and train classification model.In the on-line phase,the signal is sampled by compressed sensing.After receiving the measurements,the receiver directly extracts the feature coefficients through the feature transformation matrix,and then performs on-line classification or prediction in real time on the classification model.Experiments show that this method can achieve efficient and privacy preserving online analysis.