Robust camera model identification using demosaicing residual features

Chen Chen, Matthew C. Stamm

Abstract: In this paper, we propose a new framework for performing accurate and robust camera model identification by fully exploiting demosaicing information in a camera’s output images. Instead of fitting a camera’s demosaicing process into parametric models, our framework works by exposing and extracting a diverse set of intra-channel and inter-channel color value correlations originated from the demosaicing process. To expose these correlations, we first apply a number of diversified baseline demosaicing algorithms to re-demosaic the image under investigation, and gather a set of both linear and nonlinear demosaicing residuals. To further extract demosaicing correlations with respect to the color filter array (CFA) structure, co-occurrence matrices are calculated using a new set of geometric patterns. These patterns are specifically designed to extract different types of color value dependencies within the repeated lattice of the CFA pattern. We design a multi-class ensemble classifier to utilize all extracted color value correlations to perform camera model identification. A series of experiments show that our proposed framework can achieve an accuracy of 98.14% on a database with 68 camera models, and is highly robust to post-JPEG compression and contrast enhancement.