Blind Forensics For Detecting Change Of Low Order Statistics In Natural Image

The advent of low-cost digital imaging equipment and sophisticated image editing software has not only brought convenience,but also malicious forgery.The forgery increasingly threatens the authenticity of the news,judicial impartiality,reliability of scientific research,and rights and interests of individuals,such as right of portrait.How to detect the authenticity of image is especially important.Thus digital forensics becomes a research hotspot.Traditional digital forensic,such as watermark and digital signature,usually insert prior information into an image,extract it as evidence when this image is needed to detect.The major drawback of this approach is that it is impossible to add prior information for every digital image transmitted on Internet.Therefore,the study of blind forensics of natural image is the key to detect digital forgery.According to research on blind forensics of natural image with change of low order statistics,approaches,including building model,designing and simulating algorithm, are respectively given for re-sampling image,copy-move image and splicing one.Based on analyzing the principle,theoretical background,experiment environment, mathematical model and algorithm of current blind forensics of natural image,a general model was given.Periodic correlation of re-sampled pixels is employed to realize pattern classification.Considering the singular matrix exists in generalized EM algorithm,an improved EM algorithm is proposed in this paper to detect the trace of re-sampling.Moreover,by the analysis of copy-move forgery,DCT coefficient and PCA score was used as characteristics of image.In order to ensure the robustness and efficiency of algorithm,an approach with combination of quantization and singular value decomposition was proposed.At last,phase congruency was used as characteristics to detect for splicing image. Based on the principle of Support Vector Machine and verification of empirical parameter,a model was trained by using samples from image dataset.This model provides pattern class to predict splicing forgery.