Efficient and privacy-protected content-based image retrieval without homomorphic encryption

Images play an increasingly important role in our lives as carriers of Information. At the same time, personal devices are increasingly unable to store and compute large amounts of images, so it is necessary to outsource to cloud servers. However, this also brings privacy issues, such as personal photos, medical photos, map mapping, etc. In this paper, we present a content-based image retrieval (CBIR) scheme that can be performed on ciphertext images. In order to better represent the Lage, instead of the traditional algorithms such as SIFT and HOG, image features are extracted using fine-tuned convolutional neural networks, and then outsources the encrypted feature vector and ciphertext image to cloud server. Considering the efficiency of the search, this paper uses the k-means algorithm to and local sensitive hash function to build a secure tree index. This paper proposes a new functional encryption of the inner product to calculate the Euclidean distance between the feature vectors to obtain the similarity between the images, and finally return the ciphertext image satisfying the condition to the user. The experimental results show the efficiency of our program, in addition, the paper gives security analysis to prove that our solution can against Chosen-Plaintext Attack (CPA).