Validation of likelihood ratio methods for forensic evidence evaluation handling multimodal score distributions

This article presents a method for computing Likelihood Ratios (LR) from multimodal score distributions produced by an Automated Fingerprint Identification System (AFIS) feature extraction and comparison algorithm. The AFIS algorithm used to compare fingermarks and fingerprints was primarily developed for forensic investigation rather than for forensic evaluation. The computation of the scores is speed-optimized and performed on three different stages, each of which outputs discriminating scores of different magnitudes together forming a multimodal score distribution. It is worthy mentioning that each fingermark to fingerprint comparison performed by the AFIS algorithm results in one single similarity score (e.g. one score per comparison). The multimodal nature of the similarity scores can be typical for other biometric systems and the method proposed in this work can be applied in similar cases, where the multimodal nature in similarity scores is observed. In this work we address some of the problems related to modelling such distributions and propose solutions to issues like data sparsity, dataset shift and over-fitting. The issues mentioned affect the methods traditionally used in the situation when a multimodal nature in the similarity scores is observed (a Kernel Density Functions (KDF) was used to illustrate these issues in our case). Furthermore, the method proposed produces interpretable results in the situations when the similarity scores are sparse and traditional approaches lead to erroneous LRs of huge magnitudes.