Digital Image Passive Forensics Research Based On Binary Similarity Measures

With the rapid development of computer and multimedia technologies, the wide use of digital equipment such as cameras, printers and scanners have make digital images all over the world. The use of images brings us some problems, for instance, the authenticity of image content authentication, the integrity of image information protection. Image forgers can easily edit an image by utilizing image processing software. Frequently-used manipulating methods include splicing, combining, blurring and so forth make the manipulation hard to be detected, which brings great trouble to the domains of news, law, science and finance, and causes crisis of confidence to the internet. Digital image passive forensics can directly detect images based on content of tampering images. Therefore, the study of digital image passive forensics technology is significant to detect image manipulation.In this thesis, we analyze the principles and characteristics of present image forensics technology, study the theoretical background, experimental environment, mathematical models and algorithms of existing methods, emphasis on binary similarity measures based digital image passive forensics.In this thesis, the binary similarity measures algorithm is applied to image passive forensics. By detecting the changes of correlation between the bit planes and the binary textual characteristics within the bit planes, we classify an original image and a tampered image. Firstly, we extract the gray value of images by DT-CWT. Secondly we extract the image features based on algorithms of one-step co-occurrence values, normalized histograms and local binary patterns. Finally, we optimize the measures with Sequential Floating Forward Selection algorithm.We design an original/tampered image classifier based on support vector machine. We present the experimental model and process of passive authentication system based on binary similarity measures, and then we design and implement the system. The simulation based on classifying images with five common operations of manipulation (scaling-up, rotation, Gaussian blurring, brightness, sharpening). The results of simulation show that our system has high classification accuracy. We propose an improved method which tries to training a new classifier for the image samples within the classification margin, the results of this method show better classification performance.