Image forensics based on reverse engineering
Today with the advent of low-cost imaging devices, such as smart phones, digital cameras and surveillance video systems, digital images become quite common in our everyday life. People tend to believe the scene they have seen, even if the scene is presented in the form of an digital image, as a proverb says, 'Seeing is believing'. However are those images really trustworthy as people have thought? In this mul- timedia world, with the wide-spread availability of those sophisticated image-editing software, such as PhotoShop and Gimp, it is easy for people to modify images to hide some information or to add a non-existing scene. These manipulations usually leave no visual clues in the tampered image. As a result, the above proverb no longer holds. To address this problem, `digital image forensics' was developed. Digital image forensics aims to verify the authentication and integrity of a digital image, without the knowledge of any prior information about the questioned image. It mainly includes two tasks: to determine whether an image is authentic and to identify the source camera of an image. What distinguishes the original image from the manipulated image is the acqui- sition process inside the digital camera, which should naturally be the only reliable solution to conquer this problem. In this work, we analyze some key operations along the image acquisition pipeline, and use the cracked information to perform forensic tasks. The contributions can be grouped into three categories: white balance(WB) , color demosaicking and defocus aberration blurs. The thesis starts with exposing which white balance algorithm has been applied in the imaging pipeline. The theoretical basis lies on the fact that, given an image, applying the same white balance operation again would not change the image. With the proposed approach, the average accuracy of source camera identification is 99.3% for 5 cameras of different brands, 98.6% for 17 cameras of different models, and 98.5% for 15 cameras equally from 3 models. This is the first time white balance has been used in source camera identification, and it leads to an almost perfect result. Most commercial cameras have only one CCD/CMOS sensor, which produces just a gray scale image. In order to get a colored one, cameras apply a process called demosaicking. This thesis estimates the model and parameters of the demosaicking process to detect forgery. With this method, we can identify which part of the image that is inconsistent with the rest, in the form of their corresponding estimation error. This is the first time that the copy-move area from another image can be exposed using demosaicking. The third part of this thesis aims at integrity verification using image defocus blur. We can calculate the image defocus aberration, and estimate its depth information. Also from defocus aberration consistency, we can determine whether an image has been altered. This is the first time defocus blur has been used to perform forensic task. The proposed method increases the average accuracy of splicing detection to 81%, while the best existing published result using the same database is only 68.8%.