Encrypted Image Retrieval Scheme Based on Neural Network and LBP Mapping

The privacy-preserving image retrieval research has been a hot issue for cyber-physical system security. Image encryption could effectively meet the growing privacy needs of users. However, the existing searchable encryption schemes are computationally complex, time-consuming, and have low retrieval accuracy. What’s more, the image feature extraction requires multiple round communications between the cloud server and the image owner, which causes additional communication burdens. To solve these problems, a privacy-preserving image retrieval scheme combining deep learning technology and Local Binary Pattern (LBP) features is proposed. To protect images, block permutation, order-preserving encryption, and intra-block pixel substitution with multiple tables are used. The encrypted images would be uploaded to the cloud server and the LBP maps of encrypted images were calculated on the cloud side. Then, the neural network was trained by the encrypted LBP maps, and the image features were extracted from them with the trained neural network. The extracted features establish the image index. The Manhattan distance of features was utilized to measure the similarity between the stored images. Finally, the images similar to the query image were returned to the image owner. The security analysis and experimental results show that the proposed scheme is better than most existing schemes regarding security and retrieval accuracy.