DHAN: Encrypted JPEG image retrieval via DCT histograms-based attention networks

计算机科学 加密 离散余弦变换 人工智能 上传 JPEG格式 计算机视觉 卷积神经网络 模式识别(心理学) 图像(数学) 计算机网络 操作系统
作者
Qihua Feng,Peiya Li,Zhixun Lu,Zhibo Zhou,Yongdong Wu,Jian Weng,Feiran Huang
出处
期刊:Applied Soft Computing [Elsevier BV]
卷期号:133: 109935-109935 被引量:10
标识
DOI:10.1016/j.asoc.2022.109935
摘要

In image retrieval, the images to be retrieved are stored on remote servers. Since the images contain amounts of privacy information and the server cannot be fully trusted, people usually encrypt their images before uploading them to the server, which raises the demand for encrypted image retrieval (EIR). Current EIR techniques extract ruled hand-craft features from cipher images first and then build retrieval models (e.g., support vector machine, SVM) by these features, or use deep learning models (e.g., Convolutional Neural Network, CNN) to learn cipher-image representations in an end-to-end manner. However, SVM is not skilled at learning non-linear embedding in complex image database, and end-to-end EIR leads to low image security or retrieval performance because CNN is sensitive to extreme chaotic cipher images. Not-end-to-end EIR offers excellent encryption performance, and deep learning can further mine non-linear embedding from ruled hand-craft features. To this end, we propose a novel EIR scheme, named discrete cosine transform (DCT) Histograms-based Attention Networks (DHAN), which is based on deep learning to enhance expression ability of cipher-image in a not-end-to-end manner. Specifically, the DCT coefficients of images are encrypted by value replacement and block permutation encryption, and then the effective histogram features of DCT coefficients are extracted from the cipher images since the sets of DCT frequency in encrypted images are similar to that of plain images. After that, to dynamically learn the salient features of cipher images, we propose a new module named ResAttention and design deep attention networks to provide retrieval. Extensive experiments on two datasets demonstrate that DHAN not only provides high image security but also greatly improves retrieval performance than that of existing schemes.
最长约 10秒,即可获得该文献文件

科研通智能强力驱动
Strongly Powered by AbleSci AI
更新
PDF的下载单位、IP信息已删除 (2025-6-4)

科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
刚刚
1秒前
传奇3应助MLJ采纳,获得10
1秒前
哎呀妈呀完成签到,获得积分10
1秒前
干净海秋发布了新的文献求助10
2秒前
2秒前
4秒前
4秒前
5秒前
Evooolet完成签到,获得积分10
5秒前
ding应助凡是过往皆为序章采纳,获得10
7秒前
干净海秋完成签到,获得积分10
7秒前
小蘑菇应助zd采纳,获得10
7秒前
Ava应助别偷我增肌粉采纳,获得30
7秒前
abc完成签到,获得积分10
8秒前
一点通发布了新的文献求助10
9秒前
12秒前
淡淡的豁应助棒棒采纳,获得150
14秒前
小秋完成签到,获得积分10
14秒前
华仔应助木木林采纳,获得10
14秒前
QIZH发布了新的文献求助10
14秒前
FashionBoy应助py999采纳,获得10
15秒前
小马甲应助圆蓬蓬采纳,获得10
16秒前
16秒前
Orange应助陈念采纳,获得10
17秒前
17秒前
热心又蓝完成签到,获得积分10
18秒前
111发布了新的文献求助10
18秒前
曾经晓亦发布了新的文献求助20
18秒前
量子星尘发布了新的文献求助10
18秒前
FashionBoy应助一点通采纳,获得10
18秒前
害怕的冬云完成签到,获得积分10
19秒前
丘比特应助nni采纳,获得20
21秒前
陈一完成签到 ,获得积分10
21秒前
现代完成签到,获得积分10
22秒前
尹天奇发布了新的文献求助10
22秒前
22秒前
SYLH应助QIZH采纳,获得10
23秒前
结实的泥猴桃完成签到 ,获得积分10
24秒前
tex关闭了tex文献求助
25秒前
高分求助中
A new approach to the extrapolation of accelerated life test data 1000
Picture Books with Same-sex Parented Families: Unintentional Censorship 700
ACSM’s Guidelines for Exercise Testing and Prescription, 12th edition 500
Nucleophilic substitution in azasydnone-modified dinitroanisoles 500
不知道标题是什么 500
Indomethacinのヒトにおける経皮吸収 400
Phylogenetic study of the order Polydesmida (Myriapoda: Diplopoda) 370
热门求助领域 (近24小时)
化学 材料科学 医学 生物 工程类 有机化学 生物化学 物理 内科学 纳米技术 计算机科学 化学工程 复合材料 遗传学 基因 物理化学 催化作用 冶金 细胞生物学 免疫学
热门帖子
关注 科研通微信公众号,转发送积分 3975755
求助须知:如何正确求助?哪些是违规求助? 3520108
关于积分的说明 11200829
捐赠科研通 3256492
什么是DOI,文献DOI怎么找? 1798298
邀请新用户注册赠送积分活动 877509
科研通“疑难数据库(出版商)”最低求助积分说明 806403