Image Source Identification Based On Virtual Sample Generation

With the development of science and technology and the emergence of various intelligent devices,digital image has become a very popular media for information dissemination,which also increased the risk of digital images being tampered.In order to restrict such things by effective means,the forensics technology of digital image has developed rapidly.Digital image forensics technology can be classified into active and passive categories.The source identification based on camera model in this paper belongs to a very important branch of digital image passive forensics.At present,the research results on the direction of digital image source identification are very rich,but the premise of these methods is to have enough training samples.The accuracy of source identification will be greatly affected when the prior information of samples is insufficient.In view of this actual situation,specific solutions are given,and the main research contents are as follows:(1)Firstly,a source identification method based on virtual sample ensemble learning is proposed.Virtual sample generation includes two methods: mean method and MTD method based on trend diffusion.Mean method is to generate virtual samples by adding the average values between every two samples in the same class,then add the generated virtual samples to the original samples to form a new training sample and send them to support vector machine(SVM)for training.The MTD method firstly calculates the diffusion range of samples according to the trend diffusion theory,and virtual samples are randomly generated according to the uniform distribution in this range.Finally,the virtual samples with high probability are added to the original training set.Experiments show that the randomness in generating virtual samples leads to less improvement in the final classification results.Based on the defect of virtual sample generation method,the idea of ensemble learning is added.The number of groups of virtual samples are used to train the base classifiers.Finally,the voting method is used to integrate the learning results of each base classifier,and the final classification results are obtained.The experimental results in the same experimental environment confirm that the source identification method based on virtual sample ensemble learning is superior to the method based only on virtual sample generation.(2)A semi-supervised method is added to the ensemble learning,i.e.the source identification method based on virtual sample semi-supervised ensemble learning.Firstly,appropriate unlabeled samples are selected as new training samples according to theclassification results of virtual sample ensemble learning method,and new virtual samples are generated.The new training samples and the new virtual samples are trained by ensemble learning again,and the final classifier can be obtained.Experiments show that this method has the best effect.