Speaker: Pablo Ramírez Hereza.

Abstract: The forensic comparison of glass task aims to compare a glass sample of unknown source with a control glass sample of known source. In this work, we use multielemental features from laser ablation inductively coupled plasma (LA-ICPMS) to compute a likelihood ratio. This calculation is a complex procedure that generally requires a probabilistic model of the within-source and between-source variability of the glass sample features. The assumptino that the distribution within each measurement group (and even within source) is normally distributed is a safe one with the available data. However, the variability between sources from different glasses cannot be assume to be normally distributed. That is why we have used a kernel density distribution to describe the between-source variation. In this work, instead of modelling distributions with complex densities (such as kernels), we propose to Gaussianize the data and to use a model with Gaussian assumptions. We believe that this simplification will help to have a model less sensitive to the data dimensionality, and more prone to be extended with Bayesian techniques. However, this assumption of normality represents a strong limitation on the performance of the likelihood ratio. Thus, in this context, in order to obtain a better fit of the features with the Gaussian model assumptions, we propose the use of different normalization techniques of the LA-ICP-MS glass features, namely marginal Gaussianization based on histogram matching, and a more complex joint Gaussianization using normalization flows. We report an increase in the performance of the likelihood ratios computed by the model and, consequently, a more reliable forensic comparison model.