摘要
Copy-move是一种常用的图像伪造手段,它通过复制图像的某一区域,移动并粘贴到同一图像的其他位置,达到掩盖重要信息或伪造虚假场景的目的。近年来,为了防止copy-move被用于违法犯罪,copy-move伪造检测技术迅猛发展,在维护社会运行秩序和信息安全方面发挥着积极作用。本文提出一种基于条件生成对抗网络(conditional Generative Adversarial Networks,cGANs)的copy-move伪造检测方法。针对图像copy-move伪造检测,该方法优化设计了cGANs的损失函数,并使用适量的弱监督样本来提升网络性能。不同于目前大部分检测算法,该方法不仅可以定位出图像中的相似区域,还可以有效区分伪造来源区域和伪造目标区域。实验结果表明,本文所提出的方法在检测准确率上显著优于现有方法。
Copy-move is a common attacking method for producing image forgeries.A certain area of an image is copied and then pasted over another area of the image to conceal important information or construct a fake scene.In recent years,in order to prevent copy-move attack from being abused in illegal activities,the methods for copy-move forgery detection have been developed rapidly.These forensics methods play positive roles in maintaining social order and securing information.In this paper,based on conditional generative adversarial networks(cGANs),a novel method is proposed for the detection of copy-move forgery.To boost the detection performance,the loss function of cGANs is optimally designed,and an appropriate amount of weakly supervised samples are utilized to improve the network.Unlike most existing detection methods,the proposed method can not only detect similar regions in an image,but also effectively distinguish between source forgery regions and target forgery regions.Extensive experimental results show that the proposed method remarkably outperforms the compared methods in detection accuracy.
作者
李应灿
杨建权
丁峰
朱国普
Li Yingcan;Yang Jianquan;Ding Feng;Zhu Guopu(Shenzhen Institutes of Advanced Technology,Chinese Academy of Sciences,Shenzhen,Guangdong 518055,China;School of Software,Nanchang University,Nanchang,Jiangxi 330031,China)
出处
《信号处理》
CSCD
北大核心
2020年第9期1533-1543,共11页
Journal of Signal Processing
基金
国家自然科学基金项目(61872350,61802382,61572489)
广东特支计划(2019TQ05X696)
广东省自然科学基金项目(2020A1515010640)
深圳市基础研究项目(JCYJ20170818163403748)。
关键词
图像取证
copy-move伪造
伪造检测
篡改定位
条件生成对抗网络
image forensics
copy-move forgery
forgery detection
tampering localization
conditional generative adversarial networks