PS12 - Firearms issues
Deparment/unity chief, Forensic Departement "Ballistic", IRCGN
Firearms examiners use automated ballistic identification systems on a day to day basis in caseworks. But there has been a constant struggle to define a stopping threshold when browsing the hit list, or more generally what is a "good score".
Furthermore, these powerful systems may often be underused, when limited to basic search for a hit through a database.
We have conducted a study to merge the scores returned by Evofinder with Bayesian machine learning techniques, in order to affect a likelihood ratio of hit vs. non-hit for each candidate in the list. It would then allow to rely on a validated scale to define what would be a good score, and to ease the definition of a stopping threshold.
The algorithms have been trained on a database composed of 120 9 mm Luger bullets and cartridge cases, with different ammunition and gun brands. Evofinder scores distributions have been defined as mixtures of Gaussians or Binomials,
each component representing either the hits or the non hits distributions. These scores have been merged using Gibbs sampling to assess the probability for each candidate to be part of one or the other component.
The results show a clear improvement of the probability of finding a hit when using simultaneously two test fires from the questioned firearm, rather than only one. They also show that this new score is a little more efficient than Evofinder
default scores. Eventually, this unique and easily interpretable score can be directly used in a more general evaluation process, combined with other observations such as microscope comparisons.
It then extends the scope of this system to be used an evaluation of the comparison through a relevant reference population.
Ballistic Identification, Evidence Evaluation