Evaluating match confidence in automated face recognition via likelihood ratio determination: A case study

Forensic facial examinations (FFE) primarily relies on manual image examination by trained experts using standardized protocols to identify suspects through detailed comparisons. In contrast, automated facial recognition (AFR) employs machine learning and AI algorithms to generate match scores between faces, producing ranked lists of potential suspects within a given population. In both these approaches, final identification decisions must still be validated and justified by qualified practitioners to ensure legal accountability and court admissibility. This study presents an innovative FFE workflow based on the output of AFR and empowered by the use of a Bayesian statistical software tool based on kernel density estimation (KDE) for likelihood ratio (LR) calculation, demonstrated through a case study conducted with the Carabinieri Investigation Department. The methodology is based on (1) the generation of match scores between facial images using automated software, (2) the calculation of LRs through statistical modeling against reference population data, (3) Tippett Plot validation aligned with ENFSI (European Network of Forensic Science Institutes) guidelines to demonstrate model accuracy to ensure forensic reliability beyond mere LR calculation. The proposed framework provides court-admissible statistical results for facial comparisons while maintaining practitioner oversight. This methodological approach can support forensic practitioners in the courtroom with a statistical result related to the use of AFR tools. This methodology has been shared with ENFSI experts in facial comparison.