摘要
传统的感知哈希算法通过提取图像特定属性生成感知哈希序列,难以充分利用原始图像全部特征信息,影响了基于感知哈希的图像内容认证与版权保护能力.本文提出一种基于双向生成对抗网络(Bidirectional Generative Adversarial Network,BiGAN)的无监督感知哈希图像内容取证算法,基于编码网络、生成网络和判别网络间的双向迭代对抗,生成具有较强图像语义特征表示能力的感知哈希码;并通过在编码网络和生成网络间添加跳接层网络结构,将原始图像不同维度的特征信息传递到生成网络,提高生成网络语义特征学习能力与网络收敛速度;同时,在对抗损失中添加MSE误差损失,增强生成图像的视觉质量与细节表示能力;最后,基于网络间的多重迭代与对抗训练,输出兼具相同内容图像认证鲁棒性和不同内容图像区分敏感性的高性能图像感知哈希码.本研究首次采用大型图像数据库进行算法性能评价,实验结果表明基于双向生成对抗网络的感知哈希图像内容取证算法与当前其他优秀研究方案相比具有更强的图像内容取证性能.
The traditional perceptual hash algorithm creates image perceptual hash code by extracting image features with a pre-designed scheme.As it is hard to make full use of image inherent semantic characters,the performance of perceptual hash code on image content authentication and copyright protection is constrained.In this paper,an unsupervised perceptual hash algorithm for image forensics based on Bidirectional Generative Adversarial Network(BiGAN)is proposed.The main contributions of the paper are as follows:Firstly,depending on the bidirectional iterative adversary among the coding network,the generative network,and the discriminative network,the powerful learning ability of BiGAN on image inherent feature extraction is fully developed;so that the perceptual hash code that has strong image semantic feature representation capability can be created.As a result,both the identification robustness for images with identical content and the discrimination sensitivity for images with different contents are achieved.Hence,the capability of image forensics is improved.Secondly,a BiGAN optimization framework is constructed by adding a skip-connection structure between the coding and the generative network.By concatenating the shallow and deep layers′features of the sampled image,different dimensional features are organically integrated to improve the learning efficiency and the convergence speed of the proposed scheme.Thereby,the semantic information representation ability of the perceptual hash code is enhanced,and the identification robustness for identical content images is heightened.Thirdly,a Mean Square Error(MSE)loss-based performance optimization strategy for BiGAN is investigated.By computing the difference between the output of the coding network and the generative network,not only the visual quality of the generated image but also the representation capability of the generated perceptual hash code is effectively improved.Consequently,the discrimination sensitivity for different content images is intensified.In
作者
马宾
王一利
徐健
王春鹏
李健
周琳娜
施云庆
MA Bin;WANG Yi-Li;XU Jian;WANG Chun-Peng;LI Jian;ZHOU Lin-Na;SHI Yun-Qing(Department of Computer Science And Technology,Qilu University of Technology(Shandong Academy of Sciences),Jinan 250300;Shandong Provincial Key Laboratory of Computer Network,Jinan 250013;Department of Computer Science And Technology,Shandong University of Finance and Economics,Jinan 250014;Department of Cyber Security,Beijing University of Posts and Telecommunications,Beijing 100876;Department of Electrical and Computer Engineering,New Jersey Institute of Technology,Newark 07102,USA)
出处
《计算机学报》
EI
CAS
CSCD
北大核心
2023年第12期2551-2572,共22页
Chinese Journal of Computers
基金
国家自然科学基金(62272255,61872203)
国家重点研发计划(2021YFC3340600)
山东省自然科学基金(ZR2019BF017,ZR2020MF054)
山东省重大科技创新工程项目(2019JZZY020127,2019JZZY010132,2019JZZY010201)
山东省高等学校青创人才引育计划(SD2019-161)
济南市“高校20条”引进创新团队(2019GXRC031)
济南市“高校20条”工作室带头人(2020GXRC056)
济南市市校融合发展战略工程项目(JNSX2021030)资助。
关键词
图像取证
生成对抗网络
感知哈希
跳接
均方误差
image forensics
generative adversarial network
perceptual hash
skip-connection
mean square error
