基于对偶生成对抗网络的数字成像设备指纹提取算法  

Fingerprint Extraction Algorithm for Imaging Equipment Based on Dual Adversarial Network

作　　者：侯春晖 田华伟[1,2] 肖延辉[1,2] 郝昕泽 HOU Chunhui;TIAN Huawei;XIAO Yanhui;HAO Xinze(School of National Security,People's Public Security University of China,Beijing 100038,China;Research Center for Public Security Information,People's Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学国家安全学院,北京100038 [2]中国人民公安大学公安情报研究中心,北京100038

出　　处：《中国人民公安大学学报（自然科学版）》2021年第3期79-85,共7页Journal of People’s Public Security University of China(Science and Technology)

基　　金：中国人民公安大学公共安全行为科学实验室开放课题(2020SYS05);国家自然科学基金(61972405,61772539)。

摘　　要：光响应非均质性具有唯一性和相对稳定性,是对数字成像拍摄设备进行溯源取证的重要依据,被称为设备指纹。在图像的真实噪声中,除了PRNU设备指纹外,还包含大量的随机噪声。为有效提取PRNU设备指纹,提出一种基于对偶生成对抗网络的数字成像设备指纹提取算法。该算法基于生成对抗网络的思想进行训练,生成训练数据进行数据增强,有效地提高了训练效率;同时采用UNet的网络结构联合残差学习的策略,能够在减少GPU占用的同时减少对训练数据的需求量。提出的算法在当前最大的智能手机来源取证数据集上进行测试,实验结果显示,与基于块匹配3D滤波以及前馈去噪卷积神经网络的算法相比,该算法有着更好的识别效率和普适性。Photo-response non-uniformity(PRNU)is an important basis for traceability of imaging equipment,and is known as fingerprint of imaging equipment.In this paper,a new fingerprint extraction algorithm based on dual adversarial network for imaging equipment is proposed.Dual adversarial network is based on the idea of generative adversary network for training,and adopts the network structure of UNet combining residual learning,which can reduce the occupation of GPU and the demand for training data.At the same time,the noise generation model in this paper can generate training data of further training of the image denoising network.The algorithm in this paper has been tested in the current largest smart phone source forensics dataset.The experimental results show that the algorithm in this paper has better recognition efficiency and universality compared with the algorithm of BM3D and xDnCNN.

关 键 词：图像来源取证 对偶生成对抗网络 光响应非均质性 

分 类 号：D918.91[政治法律—法学]