Bayesian evaluation of dynamic signatures in operational conditions
Linden Jacques,Bozza Silvia,Marquis Raymond,Taroni Franco
DOI: https://doi.org/10.1016/j.forsciint.2022.111173
Abstract:Forensic handwriting examiners (FHE) activities are focused on comparative analysis of handwritten objects such as signatures. Their role is to provide and evaluate evidence for and against the authenticity of a questioned signature. In recent years, cases involving handwritten signatures captured on electronic devices have become more commonplace. These so-called 'dynamic signatures' (also known as 'digitally captured signatures') are much different from paper-based signatures. Not only does the medium of recording differ, but also the type, volume of data and features are different from the pattern-based evidence that makes up paper-based signatures. Recent developments in forensic science - including signature examination - have led to the adoption of evaluative probabilistic methodologies in many disciplines [see, e.g. ENFSI 1915 Guidelines]. In the current paper, a probabilistic model to evaluate signature evidence in the form of multivariate data, as proposed and described in Wacom Europe GmbH (2019), is adopted. Topics like data sparsity, joint evaluation of multiple features and feature selection are investigated. Performed experimental studies showed an accuracy rate above 90% even when a limited number (5) of reference signatures was available. The performances of a multivariate approach are compared with those characterizing a so-called multiplicative approach where variables (features) are taken as independent and the Bayes' factor (BF) is obtained as the product of univariate BFs associated to each selected feature. The simplicity of this latter approach is, however, accompanied by severe issues about the reliability of results. The use of a multivariate approach is therefore highly recommended. Finally, the evidential values in correspondence of alternative feature sets are compared. Results suggest that discriminative features are writer-related and necessitate a case-specific selection.