A Comparison of Similarity Scores Between Bullet Casings: Forensic Analysts Versus an Algorithm

• Activity Number: 291 - Statistical Applications in Forensic Evidence
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Advisory Committee on Forensic Science
• Author(s): Maria Cuellar*
• Companies: Carnegie Mellon University
• Keywords: firearms; ballistics; forensic; machine learning; cartridges

Abstract

The quality of firearm analysis relies on examiners' ability to identify whether two cartridges were fired from the same gun (ie. match). It is possible that analysts perform worse in in difficult tasks (matches that look dissimilar and non-matches that look similar) than in easy tasks (matches that look similar and non-matches that look dissimilar). We used an experimental dataset of high-quality images of breech faces of cartridges from the National Institute for Standards in Technology (NIST) called NBIDE. We ranked the pairs by using a similarity metric provided by a NIST algorithm that was used and enhanced by Tai et al. We surveyed novices (untrained individuals) to determine how they performed in difficult and easy tasks. We were able to compare the performance of novices to that of the algorithm in determining which pairs were matches. We found that novices were able to distinguish between matches and non matches even in the difficult scenarios. We repeated the experiment only with the dissimilar matches and similar non matches and found the same results. This implies that even novices are very good at firearm identification, regardless of their training.