Digital Camera Identification from Images – Estimating False Acceptance Probability

Photo-response non-uniformity noise present in output signals of CCD and CMOS sensors has been used as fingerprint to uniquely identify the source digital camera that took the image. The same fingerprint can establish a link between images according to their common source. In this paper, we review the state-of-the-art identification method and discuss its practical issues. In the camera identification task, when formulated as a binary hypothesis test, a decision threshold is set on correlation between image noise and modulated fingerprint. The threshold determines the probability of two kinds of possible errors: false acceptance and missed detection. We will focus on estimation of the false acceptance probability that we wish to keep very low. A straightforward approach involves testing a large number of different camera fingerprints against one image or one camera fingerprint against many images from different sources. Such sampling of the correlation probability distribution is time consuming and expensive while extrapolation of the tails of the distribution is still not reliable. A novel approach is based on cross-correlation analysis and peak-to-correlation-energy ratio.