Design of projection matrices for PRNU compression

Photo Response Non-Uniformity (PRNU) is the defacto standard in image source identification, allowing scientists, researchers, forensics investigators and courts to bind a picture under investigation to the specific camera sensor that took the shot at first place. Caused by silicon sensor imperfections, PRNU is characterized as a Gaussian i.i.d weak multiplicative noise embedded into every digital photo at acquisition time. Despite PRNU nearly-flat spectral characteristics, it undergoes several interpolations steps while image is demosaicked and optionally JPEG compressed. In this paper we propose a novel approach to the design of projection matrices tailored to PRNU compression. Joint effect of interpolation and projection on cross-correlation test is first analyzed, in order to derive those conditions that maximize detection while reducing false-alarm probability. A design methodology to build effective projection matrices is then presented, taking into account computational complexity. Validation of the proposed approach is finally performed against state-of-the-art methods on a well known public image dataset.