Research On Image Source Identification Technology Based On Camera Noise

This paper mainly studies the image source identification technology in digital image forensics.As an important research direction of digital image forensics technology,image source identification aims to determine which device take a given image,this technology is widely used in copyright protection,court forensics and other fields.This paper firstly proposes that Non-unique Artifact(NUA)noise can be used for camera model identification.Based on the analysis of existing image source identification algorithms,this paper proposes the approach for camera model classification that trains a Converlutional Neural Network model with noise residual image in a frequency domain.Aiming at the problem of image source identification in the field of digital image forensics,we study the current situation at home and abroad,summarize various features used for device model and device individual identification,and analyze the advantages and disadvantages of different features.Based on the above analysis,the innovation points of this paper are proposed:1.This paper takes NUA noise as the research object,and proposes a method for extracting NUA noise.Then the effectiveness of NUA noise for device model classification is verified through a large number of experiments.2.This paper firstly proposes a CNN-based approach for camera model classification that trains a CNN model with image noise in the frequency domain.Different from the method of training with image/noise in a spatial domain,the proposed method utilizes the periodic characteristics of NUA noise.In the experiments with 14 camera models,the classification accuracy for the proposed method is improved by about 1.5% compared with the spatial domain method.3.The existing Convolutional Neural Network technology is too dependent on training samples,but in reality,there are few image samples and many types of cameras models.In order to solve this problem,this paper introduces the Siamese network into the problem of device model classification,and proposes an improved Siamese network framework for model classification.The experiment results verify the effectiveness of the proposed method for training with few samples.Finally,this paper uses the images from the standard benchmark Dresden database and VISION database to comprehensively evaluate the performance of proposed algorithm.In the simulation of network training algorithm based on frequency domain,Alex Net reduces the training time by more than 50% compared with the more complex network Goog Le Net,and achieves more than 99% accuracy in the classification of 22 camera models and 23 mobile phone models.In the simulation of the device model classification algorithm based on the Siamese network,only one tenth of the data in the above experiment is used to achieve the classification accuracy of 86%,indicating the great potential of the Siamese network for image source classification.