基于Xception孪生结构网络的篡改图像检测方法  

Tamper Image Detection Method Based on Xception Siamese Structure Network

作　　者：童世博 孙鹏[1,2] 郎宇博 解梦达 TONG Shi-bo;SUN Peng;LANG Yu-bo;XIE Meng-da(Department of Criminal Science and Technique,Criminal Investigation Police University of China,Shenyang Liaoning 110854,China;Key Lab of Forensic Science,Ministry of Justice,Shanghai 200063,China;School of Cyberspace Security,Guangzhou University,Guangzhou Guangdong 511442,China)

机构地区：[1]中国刑事警察学院公安信息技术与情报学院,辽宁沈阳110854 [2]司法部司法鉴定重点实验室,上海200063 [3]广州大学网络空间安全学院,广东广州511442

出　　处：《计算机仿真》2024年第12期291-296,423,共7页Computer Simulation

基　　金：公安部技术研究计划项目(2020JSYJC25);司法部司法鉴定重点实验室开放课题(KF202317);辽宁省科技厅青年科技人才“育苗”项目(JYT2020130);中国刑事警察学院研究生创新能力提升项目(2022YCZD06)。

摘　　要：随着GAN等神经网络对于图像篡改的逼真程度逐步提高,篡改图像在网络中的快速传播所带来的负面影响也逐步加大。然而,现有检测方法过于追求提高检测篡改图像的精准度,忽略了训练困难,训练数据需求量大,以及当检测数据集中篡改图像和原始图像分布不均匀时出现的检测准确率大幅下降的问题。针对以上问题,提出Xception孪生结构网络的检测方法,首先使用预训练的Xcetpion网络模型组成孪生结构,选择欧氏距离作为度量距离计算公式,使用仅有少量原始图像的数据集组成图像对,通过对比损失进行反向传播加大不同图像种类图像对之间的度量距离,缩小相同种类图像对之间的度量距离。使用Xception孪生结构网络对于不同比例的原始和篡改图像测试集进行检测实验,实验结果表明,所提方法能够达到较高的检测准确率,同时在篡改图像和原始图像分布不均匀时减少准确率的下降幅度。With the gradual improvement of the fidelity of neural networks such as GAN on image tampering,the negative impact of the rapid propagation of tampered images in the network has gradually increased.However,the existing detection methods are too much in pursuit of improving the accuracy of detecting tampered images,ignoring the problems of difficult training,large demand for training data,and significant decrease in detection accuracy when the distribution of tampered images and original images in the detection dataset is uneven.In order to solve the above problems,this paper proposes a detection method for Xception Siamese structure network.Using the pre-trained Xcetpion network model to form the Siamese structure,selecting Euclidean distance as the measurement distance calculation formula,using a dataset with only a small number of original images to form image pairs,increasing the measurement distance between image pairs of different image types through backpropagation through comparative loss,and reducing the measurement distance between image pairs of the same kind.The experimental results show that the proposed method can achieve high detection accuracy and reduce the accuracy when the tampered image and the original image are unevenly distributed.

关 键 词：孪生网络 深度学习 度量学习 深度伪造 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]