Alternative Measures of Association Quality in Algorithmic Toolmark Identification

• Activity Number: 337
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Committee of Representatives to AAAS
• Author(s): Nicholas Petraco*+
• Companies: City University of New York, John Jay College of Criminal Justice
• Keywords: forensic ; toolmarks ; local false discovery rates ; conformal predition theory ; support vector machines

Abstract

Over the last several years, forensic firearm and toolmark examiners have encountered harsh criticism that there is no accepted methodology to generate numerical "proof" that independently corroborates their morphological conclusions. The focus of our research has thus been to investigate the validity of toolmark pattern analysis from an objective, algorithmic and numerical perspective; that can ultimately withstand the scrutiny of the adversarial legal system. We have successfully exploited 3D microscopy and applied various machine-learning techniques as a step towards accomplishing this goal. The most interesting research question we have from an applied statistics point of view is how to estimate the "quality" of a machine made association between a tool and a toolmark, in a falsifiable way. As a "frequentist" based approach to association quality, conformal prediction theory is used to assign orthodox confidence levels to each toolmark identification. For a "Bayesian" oriented approach we are pursuing Efron's local false discovery rate methodology. Fits are made with SVM intermediates and a modification of Storey and Tibshirani's method.