Physics-based and statistical features for image forensics = Physikalische und statistische Merkmale in der Bildforensik

The objective of blind image forensics is to determine whether an image is authentic or captured with a particular device. In contrast to other security-related fields, like watermarking, it is assumed that no supporting pattern has been embedded into the image. Thus, the only available cues for blind image forensics are either a) based on inconsistencies in expected (general) scene and camera properties or b) artifacts from particular image processing operations that were performed as part of the manipulation. In this work, we focus on the detection of image manipulations. The contributions can be grouped in two categories: techniques that exploit the statistics of forgery artifacts and methods that identify inconsistencies in high-level scene information. The two categories complement each other. The statistical approaches can be applied to the majority of digital images in batch processing. If a particular, single image should be investigated, high-level features can be used for a detailed manual investigation. Besides providing an additional, complementary testing step for an image, high-level features are also more resilient to intentional disguise of the manipulation operation. Hence, the first part of this thesis focuses on methods for the detection of statistical artifacts introduced by the manipulation process. We propose improvements to the detection of so-called copy-move forgeries. We also develop a unified, extensively evaluated pipeline for copy-move forgery detection. To benchmark different detection features within this pipeline, we create a novel framework for the controlled creation of semi-realistic forgeries. Furthermore, if the image under investigation is stored in the JPEG format, we develop an effective scheme to expose inconsistencies in the JPEG coefficients. The second part of this work aims at the verification of scene properties. Within this class of methods, we propose a preprocessing approach to assess the consistency of the illumination conditions in the scene. This algorithm makes existing work applicable to a broader range of images. The main contribution in this part is a demonstration of how illuminant color estimation can be exploited as a forensic cue. In the course of developing this method, we extensively study color constancy algorithms, which is the classical research field for estimating the color of the illumination. In this context, we investigate extensions of classical color constancy algorithms to the new field of non-uniform illumination. As part of this analysis, we create a new, highly accurate ground truth dataset and propose a new algorithm for multi-illuminant estimation based on conditional random fields.