Encrypted Image Retrieval Based On Deep Attention Networks

There are many applications of image retrieval in life,such as image search engine,similar image recommendation,face recognition and so on.In recent years,with the rapid development of cloud server and digital technology,people like to store pictures on cloud servers,which can not only break the limitation of traditional local storage space,but also can retrieve images anytime and anywhere.Image privacy security has always been focus of people’s attention,and now people are gradually realizing the importance of privacy protection.The pictures uploaded to the cloud server are no longer under the complete control of the image owner,so the images will be encrypted before uploading to the server to protect the privacy of the images.The traditional image encryption algorithm will hinder the subsequent image retrieval operations,so the encrypted image retrieval work has gradually attracted attention,and has begun to be applied and studied in some academic and practical life.Encrypted image retrieval should not only provide retrieval services,but more importantly,ensure image privacy and improve retrieval accuracy during retrieval.In order to improve the retrieval accuracy of encrypted images,we use deep learning methods to build retrieval model.Deep learning has become one of the current research hotspots in recent years.Using neural networks to build retrieval networks and learn image representations can effectively improve retrieval performance.How to combine deep neural network and encrypted image retrieval organically is also one of the key points of this paper.In this paper,we propose three new encrypted image retrieval methods.All three methods use deep attention networks to build retrieval models,which not only greatly improve retrieval performance while protecting the privacy of images.The first is encrypted JPEG image retrieval based on Huffman code and self-attention networks.In this method,we perform permutation and stream encryption on the image during the JPEG compression process,and according to the Huffman code of the encrypted image histogram constructs self-attention networks to perform retrieval.The selfattention mechanism in the retrieval model can effectively learn the global dependencies of encrypted image representations.This method can not only effectively protect image privacy,but also outperform the current encrypted image retrieval model in retrieval performance.The second is encrypted JPEG image retrieval based on discrete cosine transform(DCT)coefficients and attention networks.After the image is encrypted with value replacement and permutation encryption,we extract DCT coefficients’ histograms and build attention networks to conduct retrieval.In this retrieval model,we propose a new residual attention module for this method to learn the salient features of encrypted images and dynamically assign important weight scores to these salient features.In addition,the model combines some loss functions,regularization and post-processing techniques.The second method has higher visual security and better retrieval performance than the current encrypted image retrieval performance,and the retrieval networks of this method can be applied to the retrieval stage of other encrypted image retrieval schemes,and can significantly improve these methods’ retrieval performance.The third is the end-to-end encrypted image retrieval.Although the first two methods can effectively improve the retrieval performance,they are all based on hand-craft features extraction,such as manually extracting the DCT histogram of the encrypted image as the input of the network,which may increase the extra cost.Therefore,we propose an end-to-end encrypted image retrieval method.This method directly uses the cipher-images as input,and the retrieval model can automatically extracts features without manually extracting features.The retrieval model adopts the Vision Transformer(ViT)structure.The experimental results show that ViT can learn the representation of encrypted images better than convolutional neural network(CNN),because ViT can pay more attention to the global relationship,and CNN pays more attentions to local features,but The local features of the encrypted image have been scrambled and destroyed globally.Finally,we summarize the three encrypted image retrieval schemes proposed in this paper,and compare the advantages and disadvantages of these three methods.In addition,we propose several areas for improvement in future work,as well as encrypted image retrieval work that we will explore in the future.