Automated face recognition in forensic science: Review and perspectives

With recent technological innovations, the multiplication of captured images of criminal events has brought the comparison of faces to the forefront of the judicial scene. Forensic face recognition has become a ubiquitous tool to guide investigations, gather intelligence and provide evidence in court. However, its reliability in court still suffers from the lack of methodological standardization and empirical validation, notably when using automatic systems, which compare images and generate a matching score. Although the use of such systems increases drastically, it still requires more empirical studies based on adequate forensic data (surveillance footages and identity documents) to become a reliable method to present evidence in court. In this paper, we propose a review of the literature leading to the establishment of a methodological workflow to develop a score-based likelihood-ratio computation model using a Bayesian framework. Different approaches are proposed in the literature regarding the within-source and between-source variability distributions modelling. Depending on the data available, the modelling approach can be specific to the case or generic. Generic approaches allow interpreting the score without any available images of the suspect. Such model is henceforth harder to defend in court because the results are not anchored to the suspect. To make sure the computed score-based LR is robust, we must assess the performance of the model with two main characteristics: the discriminating power and the calibration state of the model. We hence describe the main metrics (Equal Error Rate and Cost of log likelihood-ratio), and graphical representations (Tippett plots, Detection Error Trade-off plot and Empirical Cross-Entropy plot) used to quantify and visualize the performance characteristics