Assessment of signature handwriting evidence via score-based likelihood ratio based on comparative measurement of...

This paper extends on previous research on the extraction and statistical analysis on relevant dynamic features (width, grayscale and radian combined with writing sequence information) in forensic handwriting examinations. In this paper, a larger signature database was gathered, including genuine signatures, freehand imitation signatures, random forgeries and tracing imitation signatures, which are often encountered in casework. After applying Principle Component Analysis (PCA) of the variables describing the proximity between specimens, a two-dimensional kernel density estimation was used to describe the variability of within-genuine comparisons and genuine-forgery comparisons. We show that the overlap between the within-genuine comparisons and the genuine-forgery comparisons depends on the imitated writer and on the forger as well. Then, in order to simulate casework conditions, cases were simulated by random sampling based on the collected signature dataset. Three-dimensional normal density estimation was used to estimate the numerator and denominator probability distribution used to compute a likelihood ratio (LR). The comparisons between the performance of the systems in SigComp2011 (based on static features) and the method presented in this paper (based on relevant dynamic features) showed that relevant dynamic features are better than static features in terms of accuracy, false acceptance rate, false rejection rate and calibration of likelihood ratios.