Image Data Publishing Method Based On Differential Privacy

With the rapid improvement of information level in the era of big data,the importance of image data privacy protection is gradually emerging.As an important information medium,pictures have many important properties.In particular,the medical facial data of patients,including patients' gender,age,facial expression and other data attributes,can easily lead to personal privacy leakage if these data are directly released to third-party applications for analysis.The privacy protection of image data has attracted extensive attention of researchers.How to use the differential privacy method to protect the privacy of image data and make the image data highly available has become the main research goal of scholars.In this paper,the face image data is taken as the object,and the Laplace mechanism of differential privacy is taken as the basis.The frequency domain method and algebraic method are studied respectively.The main research contents are as follows Based on the method of feature extraction in frequency domain,a WIP method of differential privacy image publishing based on wavelet transform is proposed.Wavelet transform is used to transform and compress the image,and then noise is added to the main features after transformation to obtain the published image satisfying differential privacy,so as to solve the problem of low availability of largescale image and the problem that Fourier transform can't process the mutation signal.The experimental results show that compared with the similar methods in frequency domain,the proposed method has higher recognition and information entropy,and the accuracy of PCA+SVM is 10% higher.Compared with other differential privacy image publishing methods based on frequency domain,the proposed WIP method has higher availability and robustness.Based on the algebraic feature extraction method,NMM method based on NMF,NSRA method and NESRA method based on NMF and SVD are proposed.The image is compressed by the method of algebraic matrix decomposition,the feature information of the image is extracted,the noise is added to the feature information matrix,and then the published image satisfying the differential privacy is reconstructed to solve the problems of high global sensitivity and low availability of SVD due to the change of external environment.The experimental results show that,compared with the similar methods under the algebraic method,the proposed method can obtain a higher recognition degree of the noisy image,and the information entropy is closer to the original image.The classification accuracy of NSRA and NESRA in PCA+SVM is close to or equal to that of the original image.The classification accuracy of NMM is about 5% higher than that of SRA,while that of NSRA and NESRA is about 30% higher than that of SRA,and about 5% to 10%higher than that of ESRA.The NMM method under the algebraic method proposed in this paper has higher availability,while the NSRA method and NESRA method have higher availability and robustness in similar algorithms.