Secure Image Retrieval Based On Deep CNN Features

In cloud computing era,secure image retrieval has become an important application in image database,and it has been a very active research area in multimedia and cloud computing over the last few years.The protection of image features which are the core of secure image retrieval is one of the keys to prevent user information leakage.Image features such as HOG,SIFT,CNN,etc.,are once considered to be confidential.However,recent studies have shown that these features contain ambiguous semantic information and can be reconstructed to images.Moreover,the reconstructed image has a very high visual similarity to the original image.Due to the potential immensurable information disclosure,features are not be confidential any more.Besides,there are few researches focus on feature protection and all of them encounter various inherent limitations,which lead to fragile security and low accuracy.Therefore,effectively protecting the image features is required and a severe privacy challenging in cloud computing era while subcontracting confidential retrieval tasks over massive sensitive images.Based on the introduction of cloud data leakage and image reconstruction based on image features,this thesis systematically expounds the relevant content and research status of secure image retrieval.In this thesis,we focus on relieving the risk of CNN feature's leakage while maintaining the functionality of CNN feature-based image retrieval.In addition,the research on large-scale secure image retrieval is carried out for CNN feature protection.The main innovative achievements of this thesis are as follows:(1)An image retrieval algorithm based on CNN classification layer features is proposed.By extracting the CNN classification layer feature as a global descriptor representing the image,an end-to-end retrieval method can be realized,which avoids the computational cost and error caused by other additional operations.At the same time,the CNN classification layer feature as a probability vector contains ambiguous semantic information,which can solve the problem of semantic gap.The experimental results show that the method can effectively improve the CNN features in the image retrieval precision.(2)To provide a thorough data protection mechanism for both images and features,a largescale secure image retrieval framework is carefully designed based on SANN.This method encrypts the features and images outsourced to the cloud separately,and sends the query request to the cloud through the security trap door without revealing the query object itself,thereby effectively protecting the CNN features and images.This method achieves efficient image retrieval on encrypted data by obtaining the encryption candidate set related to the query trap door and computing the k-nearest neighbor result after decryption on the client.Based on the proposed security framework,this thesis proposes a large-scale secure image retrieval algorithm based on RS-SANN and a large-scale secure image retrieval algorithm based on SANNp by using two different SANN methods.Furthermore,we analyze the security of CNN features and the framework to ensure the security of the algorithm,and carry out rigorous theoretical analysis on the accuracy and complexity of the proposed algorithm to ensure high precision and high efficiency.Finally,we do massive experimental studies on two large-scale realistic image databases to state the compatibility for different features of various CNN models and SANN mechanism.Also,the proposing secure framework can response a task of image retrieval as quickly as original CNN models with nearly identical precision.