Statistical and geometric methods for passive-blind image forensics

Passive-blind image forensics (PBIF) refers to passive ways for evaluating image authenticity and detecting fake images. This dissertation proposes a physics-based approach for PBIF, with our definition of image authenticity derived from the image generative process comprising the 3D scene and the image acquisition device. We propose one statistical method and two geometric methods for capturing the image authenticity properties and addressing three separate problems in PBIF, i.e., detecting spliced images, distinguishing photographic images from photorealistic computer graphics, and estimating camera response function (CRF) from a single image. For image splicing detection, we show a statistical method for capturing the optical low-pass property of cameras. Through analysis on a proposed model of image splicing, we can explain the bicoherence response to image splicing better than the conventional quadratic phase coupling theory. Furthermore, we propose incorporating image-content-related features to improve the performance of image splicing detection. For distinguishing photographic images from photorealistic computer graphics, we propose a geometric method for capturing the properties of the object geometry, the object surface reflectance, and the CRF. The resulting geometry feature not only provides an intuitive understanding on how photographic images are different from photorealistic computer graphics, it also classifies the two types of images better than the wavelet characteristic feature and the features derived from modeling general computer graphics. For the work on CRF estimation, we propose a geometric method based on geometric invariants for estimating CRF from a single-color-channel image. We provide an extensive analysis of the method and also propose a generalized gamma curve CRF model.