Implementation of the likelihood ratio framework for camera identification based on sensor noise patterns

Abstract

It is investigated how implementing the likelihood ratio (LR) framework works out in the case of camera identification based on image-sensor-specific noise patterns. Two typical case scenarios are considered, one with images of low quality andthe other with images of high quality. In both cases, it is possible to obtain statistical distributions having a good fit with the reference data both for ‘matching’ and for ‘non-matching’ comparisons, and LRs are determined. It turns out that if the reference data are well separated, in the case of ‘matching’ images/cameras, the statistical fit of the distribution for ‘non-matches’ is constantly evaluated in a range where there is a lack of reference data. Because of this extrapolation issue, the LRsthat emerge are not reliable. This is not a problem that is unique to camera identification: if the informative value of any forensic comparison is high the problem emerges. An alternative approach is presented which consists of choosing a threshold value separating ‘matches’ from ‘non-matches’ and quantifying the strength of evidence of being larger/smaller than this value. If sample sizes of reference data increase LR results will increase as well, and it is shown that this approach is stable.