Development of a score-to-likelihood ratio model for facial recognition using authentic criminalistic data

Automated face matching systems have emerged as a useful tool for identification purposes in criminal investigations. In a forensic context it is desirable to evaluate the findings from such comparisons as probabilities in terms of a likelihood ratio. When comparing two biometric samples, many facial recognition systems produce a score value as the output. The score describes the relative similarity between the two facial images. To obtain the likelihood ratio, it is necessary to construct a statistical model for score-to-likelihood ratio conversion. The model is highly dependent on the available training data and ideally it should reflect the relevant population as closely as possible. In order to construct a general model applicable on a national level, we use data from a national mugshot database as training data. In a full crossmatch drawing from 51563 records, we develop and evaluate five different models in a Bayesian statistical framework using a total of 9000 facial comparisons with equal distribution between same source and different source scores.