Research On The Technology Of CNN-Based Device Source Identification And The Construction Of Standard Evaluation Dataset

Smartphone capable of taking pictures have been become an essential communication and entertainment tool for everyone.Every day,a large number of mobile phone images are generated and spread.Mobile phone image acquisition is more and more convience.This not only meet people's daily needs,but also criminals are able to use their mobile phones to capture,modify,and transmit illegal images(including terrorism,pornography,abuse,etc.).When detect those cases,public security organs need to determine the source of images.In order to clarify the thinking of investigation and determine the direction of detection.Exiting image source identification technology determine the source of the image based on the information that comes with the image,but this kind of information is easy to be tampered.At present,the most popular image source identification technology is based on Photo-Response Non-Uniformity(PRNU).However,PRNU is difficult to extract and purify.The existing PRNU extraction algorithms has obvious disadvantages,which limits the improvement of the identification accuracy.Moreover,the lack of a standard evaluation dataset dedicated to the mobile phone image source identification limits the development of mobile phone image source identification technology to a certain extent.Building a standard evaluation dataset for image source identification of mobile phone and improving the performance of the image source identification technology of mobile phone have been extensively studied by academia.This paper aims at solve the problem of the lack of standard evaluation dataset for mobile phone images.Establish a mobile phone image database as a standard evaluation dataset;the evaluation dataset contains the current mainstream mobile phone brands and models,including:HUAWEI,APPLE,OPPO,VIVO and Xiaomi,there are 95 imaging devices in total,and there are up to 13 imaging devices of the same model,with about 45811 images taken.According to research needs,the image contains eight different shooting scenes and three different shooting angles to make it more in line with the actual shooting scene.The standard evaluation dataset also admitted 1480 original videos;those are used for video source identification.According to research,the dataset is the largest standard evaluation dataset for image and video source identification.In addition,in order to make the standard evaluation dataset better meet the actual needs of public security intelligence work,9700 original images taken from 57 smartphones in daily life were also collected.These daily life images are relatively private and are used as a non-public subsidiary evaluation dataset of the standard evaluation dataset.In order to improve the performance of the image source identification technology of mobile phone,this paper proposes an image source identification algorithm based on Convolutional Neural Network(CNN).First,use CNN for feature extraction,iterative scaling,reconstruction,and enhancement to obtain image noise residuals to estimate image PRNU noise;then,calculate the correlation between the extracted PRNU noise and other image noise residuals to identify the source of the image.Finally,compare the experiments on the standard evaluation dataset established in this paper.Experimental comparison results show that the proposed algorithm has a certain improvement in performance compared with similar algorithms.For the use of mobile phone images to carry out illegal and criminal acts,the algorithm can help public security organs to investigate and solve cases,and to conduct line-expanding and parallel investigations based on the identification of homologous images.