A forensic investigator performing source identification on a questioned image from a crime aims to identify the unknown camera that acquired the image. On the camera sensor, minute spatial variations in intensities between pixels, called photo response non-uniformity (PRNU), provide a unique and persistent artifact appearing in every image acquired by the digital camera. This camera fingerprint is used to produce a score between the questioned image and an unknown camera using a court-approved camera identification algorithm. The score is compared to a fixed threshold to determine a match or no match. Error rates for the court-approved camera-identification PRNU algorithm were established on a very large set of image data, making no distinction between images with different brightness levels. Camera exposure settings and in-camera processing strive to produce a visually pleasing image, but images that are too dark or too bright are not uncommon. While prior work has shown that exposure settings can impact the accuracy of the court-approved algorithm, these settings are often unreliable in the image metadata. In this work, we apply the court-approved PRNU algorithm to a large data set where images are assigned a brightness level as a proxy for exposure settings using a novel classification method and then analyze error rates. We find statistically significant differences between error rates for nominal images and for images labeled dark or bright. Our result suggests that in court, the error rate of the PRNU algorithm for a questioned image may be more accurately characterized when considering the image brightness.
Read full abstract