Research On Visual Security Index

Images are currently extremely easy to capture and distribute nowadays because of the pervasive use of smartphones and availability of network access,and they are increasingly being used for business and daily communications.Under these circumstances,the security of images is of great importance for protecting both commercial confidentiality and individual privacy.With the development in recent decades of various efficient image encryption algorithms,such as selective encryption,measurement of visual security is of fundamental importance for encrypted images,especially for selective encrypted images.Visual security index(VSI)is a commonly used primitive in the security analysis of selective encrypted images.Many VSIs,including image quality assessments(IQAs)and VSIs,have been proposed in recent years for the security evaluation of encrypted images,but they often exhibit undesirable behaviors on selectively encrypted images of low quality.Additionally,existing VSIs usually focus on comparing the similarity between plain and encrypted images,and seldom consider the definition of visual security or what visual security means.Additionally,although many IQAs have been proposed,they usually calculate the precise change of each distortion between reference and distorted images,and fail to consider the fact that not all distortions are distinguishable by human visual system(HVS).The main work and achievements of this dissertation includes:(1)Existing image encryption algorithms and security evaluation methods have been analyzed,as well as the drawbacks of existing visual security indexes.In this dissertation,existing security evaluation methods are divided into traditional security evaluation methods and visual security evaluation methods.It has been shown that the traditional security evaluation methods cannot effectively evaluate the divulgence of visual information,and the performance of existing visual security indexes are unsatisfactory on encrypted images.(2)The system model of VSI has been theoretically analyzed and the definitions of visual security and visual security index have been presented.In order to accurately measure the divulgence of visual information,taking the functionality of the human visual capability into consideration,different levels of computing power are defined by analyzing the semantic gap.Under the given threat model,the definitions of visual security and visual security index are formalized after defining the advantage of adversary.(3)Considering the functionality of human visual system,a visual security index(VSI)has been presented based on edge and texture similarities.The proposed VSI evaluates edge similarity and texture similarity between plain and encrypted images.The two acpects are computed via multi-threshold edge detection and the co-occurrence matrix.Then,through adaptive similarity weighting,these two components are integrated to obtain the proposed VSI.Experimental results demonstrate that compared with many existing state-of-the-art visual security metrics,the proposed VSI exhibits a better performance and is stability on low-quality images.(4)Considering the diversity of encryption algorithms,a perceptual visual security index(PVSI)has been presented based on keypoint matching in order to accurately evaluate the security of encrypted images under the visual security model.The proposed PVSI utilizes a feature descriptor to obtain the descriptions of keypoints as semantic content set and then evaluates SURF content similarity between plain and encrypted images.Experimental results demonstrate that compared with many existing state-of-the-art visual security metrics,the proposed PVSI exhibits better performance on specific encrypted images and publicly available image databases.(5)A novel full-reference IQA scheme has been proposed based on multi-scale fuzzy gradient similarity deviation(MFGSD)to reduce the negative impact of imperceptible distortions.Fuzzy inference system and the standard deviation of fuzzy gradient similarity are introduced to measure the quality distinction between reference and distorted images.Extensive experiments are conducted on two publicly available image databases,and compared with many existing state-of-art IQA schemes,the proposed MFGSD has better performance on different distortion types and strengths.