Research On Source Camera Forensic Based On Deep Learning

With the rapid development of image acquisition technology and the rapid popularization of digital cameras,digital images have been integrated into people's daily work and life.However,more and more image information security problems have been exposed subsequently:criminals use various simple and powerful image processing software to tamper with images to achieve the purpose of illegal crime,and illegally make use of other people's works for profit and so on..Therefore,the research of digital image forensics technology is of great significance.The camera source forensics technique is an important branch of digital image forensics.This technology can associate digital images with the devices they capture when the source of the image is unknown.Camera source forensics technology has important applications in copyright protection,image tracing,forensic identification and so on.This paper first introduces the history and research background of digital image forensics and camera source forensics,introduces the imaging pipeline and sensor pattern noise of digital cameras,analyzes the existing camera source forensics algorithm,and according to the different application scenarios of practice after done the following three work:(1)This paper presents a large-scale image set camera source recognition algorithm based on residual network.The traditional method is to design features manually and the number of classifiable devices is small(less than 20).In this paper,the deep residual network is used instead of traditional extraction mode noise,which can be used to extract the fusion features of equipment classification,The high-pass filter is designed to filter the low-frequency image content at the input end of the network.Extraction of pattern noise.We collected large-scale image datasets of up to 65 models with a total of 119 devices.The algorithm achieved an average classification accuracy of 87%at the model level and 83%at the individual level on the test set.(2)For another application scenario in practice,a Siamese Network-based camera source matching algorithm is proposed.When we only query the image and there is no information of the active camera,we can use this algorithm to match the image in two pairs to determine whether the image pair is taken by the same camera.The algorithm is improved on the basis of the Siamese Network architecture.The dual-channel feature extractor is replaced by the feature extractor proposed in the first work of this paper.The image is trained in pairs.The same camera takes the image tag as 0,and different cameras takes the image tag as 1.The algorithm gives the prediction results by matching the input images in two pairs,and determines whether it is taken by the same camera.The accuracy rate of 89%is achieved on the test set.(3)The camera source forensics system based on B-S architecture is designed and implemented.The two algorithms proposed in this paper are modular encapsulated.By deploying two algorithm modules online,a multi-dimensional camera source forensic analysis system is constructed.Through the model level and individual level camera source classification and camera source matching,the user can provide multi-angle camera source forensics analysis function,which converts complex algorithms into web pages,which is a simple and easy-to-operate system,so that more users can use our The algorithm plays a big role in the application and promotion of the algorithm.