应用注意力机制的文档图像篡改与脱敏定位  

Document image forgery localization and desensitization localization using the attention mechanism

作　　者：郑铿涛 李斌[1,2] 曾锦华 ZHENG Kengtao;LI Bin;ZENG Jinhua(Guangdong Key Lab of Intelligent Information Processing,Shenzhen Key Laboratory of Media Security,Shenzhen University,Shenzhen 518060,China;Shenzhen Institute of Artificial Intelligence and Robotics for Society,Shenzhen 518129,China;Academy of Forensic Science,Shanghai 200063,China)

机构地区：[1]深圳大学广东省智能信息处理重点实验室,深圳市媒体信息内容安全重点实验室,广东深圳518060 [2]深圳市人工智能与机器人研究院,广东深圳518129 [3]司法鉴定科学研究院,上海200063

出　　处：《西安电子科技大学学报》2023年第6期207-218,共12页Journal of Xidian University

基　　金：国家自然科学基金(U22B2047,61872244,62272314);广东省基础与应用基础研究基金杰出青年基金项目(2019B151502001);深圳市研究与发展计划项目(JCYJ2020010905008228);上海市科委技术标准项目(21DZ2200100)。

摘　　要：诸如合同、证明文件和通知书等一些重要的文档材料,常常以电子图像格式被存储和传播。然而,由于包含关键的文字信息,此类图像往往容易被非法篡改利用,造成严重的社会影响和危害;与此同时,考虑到个人的隐私安全问题,人们往往也会对这类图像做脱除敏感信息处理。恶意篡改与脱敏均会给原始图像引入额外痕迹,但在动机上存在区别,且在操作方式上也存在一定差异。因此,有必要对二者进行区分,从而更准确地定位出篡改区域。针对这个问题,提出了一个卷积编解码网络,通过U形连接获取编码器多级特征,有效学习篡改和脱敏处理痕迹;同时,在解码网络引入多个挤压激励注意力机制模块,抑制图像内容,关注更微弱的处理痕迹,提高网络的检测能力。为了有效地辅助网络训练,构建了一个包含常见篡改操作和脱敏操作的文档图像取证数据集。实验结果表明,算法模型在此数据集上表现良好,在公开的篡改数据集上也有不错的性能,并优于对比算法。同时,所提的算法对几种常见的后处理操作具有较好的鲁棒性。Some important documents such as contracts,certificates and notifications are often stored and disseminated in a digital format.However,due to the inclusion of key text information,such images are often easily illegally tampered with and used,causing serious social impact and harm.Meanwhile,taking personal privacy and security into account,people also tend to remove sensitive information from these digital documents.Malicious tampering and desensitization can both introduce extra traces to the original images,but there are differences in motivation and operations.Therefore,it is necessary to differentiate them to locate the tamper areas more accurately.To address this issue,we propose a convolutional encoder-decoder network,which has multi-level features of the encoder through U-Net connection,effectively learning tampering and desensitization traces.At the same time,several Squeeze-and-Excitation attention mechanism modules are introduced in the decoder to suppress image content and focus on weaker operation traces,to improve the detection ability of the network.To effectively assist network training,we build a document image forensics dataset containing common tampering and desensitization operations.Experimental results show that our model performs effectively both on this dataset and on the public tamper datasets,and outperforms comparison algorithms.At the same time,the proposed method is robust to several common post-processing operations.

关 键 词：文档图像 篡改定位 脱敏定位 U-Net 挤压激励注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]