Remote sensing image security retrieval method and system in cloud environment

A remote sensing image security retrieval method in cloud environment, comprising:

• utilizing, by an image owner, a pre-trained convolutional neural network (CNN) model to extract features from outsourced images as an image set, generating an encrypted searchable index based on spectral hashing with spectral rotation for dimensionality reduction and asymmetric scalar-product-preserving encryption (ASPE), carrying out prediction error marking and pixel rearrangement to create space for watermark embedding, encrypting the image set after the prediction error marking and the pixel rearrangement to obtain an encrypted image set and constructing a Merkle tree, transmitting the encrypted image set and the encrypted searchable index to a cloud server, and sending the Merkle tree to a certification authority;
• wherein the generating an encrypted searchable index based on spectral hashing with spectral rotation for dimensionality reduction and ASPE comprises:
• extracting, by the image owner, a set of feature vectors from the image set by using the pre-trained CNN model;
• performing dimensionality reduction on the set of feature vectors via the spectral hashing with spectral rotation to obtain a set of dimension-reduced feature vectors, and performing K-means++clustering on the set of dimension-reduced feature vectors to obtain a set of cluster vectors;
• and
• expanding each cluster vector in the set of cluster vectors, followed by splitting into two parts, and encrypting the two parts obtained by splitting respectively according to two random matrices;
• wherein the carrying out prediction error marking and pixel rearrangement to create space for watermark embedding and the encrypting the image set after the prediction error marking and the pixel rearrangement to obtain an encrypted image set comprise:
• performing the prediction error marking and the pixel rearrangement on each image in the image set to create a watermark embedding space;
• generating a pseudo-random matrix by a pseudo-random generator, and binarizing a current pixel in each image, followed by performing bitwise XOR encryption;
• and
• extracting starting position information for data embedding in each image, and then embedding a watermark containing copyright information into the watermark embedding space vacated by a corresponding encrypted image in the encrypted image set in a form of pixel replacement after a starting position;
• utilizing, by a query user, the pre-trained CNN model to extract features from a query image, generating a query trapdoor based on spectral hashing with spectral rotation for dimensionality reduction and ASPE, transmitting the query trapdoor to the cloud server to obtain a search result, and subsequently verifying authenticity of the search result through the certification authority;
• storing, by the cloud server, the outsourced images and the encrypted searchable index of the image owner, retrieving encrypted images according to the query trapdoor provided by the query user, returning a Top-k encrypted image structurally similar to the query image, and then embedding specific watermarks generated by the certification authority for each user into a reserved watermark embedding space of the Top-k encrypted image;
• and
• performing, by the certification authority, both user authentication and retrieval result correctness verification, realizing the user authentication by extracting and comparing specific watermark information in a suspicious image, and realizing the retrieval result correctness verification by reconstructing the Merkle tree and comparing a root hash stored by the image owner.