Research On Privacy-sensitive Content-based Image Retrieval In The Cloud

Due to the rapid growth of digital images,content-based image retrieval has become a powerful tool to manage large-scale image databases.With the prosperity of cloud computing,it is possible for a small or medium-sized enterprise to build and maintain a large-scale,cost-efficient image retrieval system on cloud platforms.Although the cloud platforms provide convenient services of storage,computation,and communication,they bring new privacy issues.Privacy-preserving image retrieval(PPIR),searching over encrypted images,is a promising technique to protect users' privacy and several PPIR schemes have been widely studied in the past decade.Although several existing schemes have been proposed to protect users' privacy to a certain degree,they still have some drawbacks.First,in practice,only some parts of the outsourced images are private,such as faces and saliency objects,so encrypting whole images would require a heavy computational burden and make further processing difficult(e.g.,scene recognition).Second,the sensitive levels of different users vary significantly,which is ignored in existing PPIR schemes.Inspired by the above limitations,this thesis proposes Sens IR,a privacy-sensitive image retrieval scheme,which intends to protect the real private regions of outsourced images as well as guarantee the retrieval accuracy.Similar to existing region-of-interest based schemes,this thesis assumes that the salient regions are private and the users' sensitive levels are consistent with the salient scores.On the basis of the assumption,(1)this thesis proposes a privacy region detection algorithm,PRDet,to optimize the private regions generated by existing saliency detection algorithm.(2)This thesis also optimizes the CNN-based image retrieval scheme using partial convolutions to reduce the impact of pseudorandom encrypted pixels.(3)This thesis notices that the retrieval accuracy would be low when the private regions are large.To address this problem,this thesis proposes an image reranking algorithm that utilizes fixed-length similarity-preserving hash to improve the accuracy of the query processing.(4)Extensive experiments are conducted to illustrate the efficiency of privacy protection and the superiority of the proposed scheme.