Research On Encrypted Image Retrieval For Privacy Protection

In the era of big data and cloud computing,users encrypt and upload sensitive images to the cloud for storage and retrieval,which facilitates data access while protecting personal privacy.However,the cloud server cannot directly search the encrypted image.The ciphertext image retrieval technology utilizes the similarity of image feature to search the encrypted image with similar plaintext content,which protects the image content and implement search.Therefore,the contradiction between image encryption and effective retrieval is solved.However,the current technology leaks image features or feature neighbor relationships,does not support the image format of compression,and multi-user scenarios.In this dissertation,the retrieval of the encrypted image is in-depth studied,including JPEG image content encryption for retrieval,image feature encryption for retrieval,and privacy protection of efficient retrieval.The main results are as follows.1.Encrypted JPEG image retrievalA retrieval scheme of the encrypted image that compatible with JPEG format is proposed.Since scrambling and Huffman coding does not affect the neighbor relations of the DCT coefficients,the retrieval of encrypted JPEG images is still feasible without leakage of the plaintext feature.In this scheme,the content owner uses the key to scramble the DCT coefficients,then re-Huffman codes the scrambled DCT coefficients,after that,the encrypted JPEG image is compatible with the JPEG format and the ciphertext of image library is uploaded to the cloud server.The authorized user uses the same key and method to encrypt the query image,then submits the encrypted JPEG query image to the cloud server.The cloud server extracts the Huffman-code histogram of the encrypted library image and the encrypted query image,respectively,then compares the similarity of the Huffman-code histogram to find the library images similar to the query image.The search result is returned to the authorized user.The scrambling changes the order of the DCT coefficients but does not change their relationship,then the scrambled DCT coefficients of nearest neighbor are encoded as new same Huffman code.Thus,the statistical Huffman-code histogram maintains the neighbor relationship of the DCT coefficients,which can perform effective retrieval of plaintext content.The advantage of the scheme is that the encrypted image is compatible with the JPEG format,and the Huffman-code histogram is encrypted and still supports effective retrieval.2.Encrypted feature retrieval with multi-key mechanismA ciphertext retrieval scheme of image feature that supports multi-user scenarios is proposed.Different users utilize different keys to encrypt the image feature for retrieval,which solves the problem of feature leakage caused by multi-user shared key.In this scheme,the content owner separately encrypts the plaintext image and its feature.The ciphertext image and encrypted feature are uploaded in pairs to the cloud server.Specifically,the plaintext feature is expanded by replication then encrypted by a content key as an encrypted feature.The authorized user extracts the query feature from a query image and uses a privacy query key to encrypt it.The encrypted query is submitted to the cloud server.The cloud server compares the inner products between the encrypted query and different encrypted features.The neighbor encrypted features of the encrypted query are found and the corresponding encrypted images are returned to the authorized user.Since the content key and query key are respectively a random matrix and its inverse submatrix,the encryption changes the neighbor relationship between plaintext features but does not change the neighbor relationship between the query feature and plaintext feature,which can perform effective retrieval.In order to make full use of the redundancy generated by feature replication,global optimization or Gaussian distribution is used to generate multiple independent inverse submatrices,which is a multi-key secure inner product algorithm.The advantage of the scheme is that the relation between plaintext features is protected and the confidentiality of the query feature in multi-user scenarios is guarantee.3.Efficient encrypted image retrieval for privacy protectionAn efficient encrypted image retrieval scheme for protecting user queries is proposed.Using the difference in the similarity of image features,the user queries different ciphertexts of the same image by different features,which avoid the problem that a cloud server can infer the query content by the frequency of request or result.In this scheme,the content owner copies the image database,encrypts the same image with different keys,and sets integrated features of different weights,then generates splitting vector from top-down clustering the integrated features,finally constructs a multi-feature index tree of high category entropy.The leaf node points to the encrypted image.The secure inner product algorithm is used to encrypt the splitting vector,so the ciphertext index tree with the encrypted image is uploaded to the cloud server.The authorized user integrates features of a query image,encrypts the query feature by the secure inner product algorithm,and submits the ciphertext query to the cloud server.According to the inner product between the ciphertext query and the ciphertext splitting vector,the cloud server searches the ciphertext index tree and returns the encrypted image to the authorized user.In order to achieve the logarithmic retrieval efficiency,a balanced binary index tree is constructed by perfect matching in the bipartite graph.In order to protect the query content,the authorized user submits query features of different weights,so the path and result that the cloud server searches the same query image are not unique,but search result is similar after decryption that achieves the same purpose.The advantage of the scheme is that the balanced index tree realizes the logarithmic retrieval efficiency,and the differentiated search path protects the query content.In summary,this dissertation studies image encryption for retrieval,feature encryption for retrieval,and privacy protection of efficient retrieval that increase data redundancy,such as file size,feature-length or database size,to resolve the contradiction between encryption and retrieval.The retrieval privacy is protected from three levels: the content of image and feature,feature relationship,and relationship query,which promotes a new model of multimedia ciphertext service.