3D printed firearm identification: A comparison of machine learning models.

Abstract

The production of three-dimensional (3D) printed firearms is concerning as these objects are unregulated and untraceable due to the lack of serial numbers. Criminals can obtain 3D models online or design their own versions using computer-aided design (CAD) software. While prior research and law enforcement efforts have focused primarily on analyzing tangible printed objects, investigations analyzing digital evidence are currently limited. This study presents a proof-of-concept approach for classifying firearm and non-firearm objects using geometric information extracted from g-code files, which serve as the executable instructions for 3D printers. It uses two feature extraction methods: the direct g-code method and the mesh construction method. The direct g-code method extracts the features as values directly from the g-code file, while the mesh construction method converts the coordinates from the g-code file into a 3D mesh, then extracts the vertices and edges from each model. We use machine learning classifiers such as random forest (RF), support vector machine, decision tree, and a convolutional neural network to classify our objects into firearm and non-firearm objects. We then apply a 10-fold cross validation on our data to validate its accuracy. The results demonstrated that the RF model, in conjunction with the mesh construction method, achieves the highest classification accuracy of 95.80%. The mesh construction method consistently outperforms the direct g-code method accuracy results, and these performance differences are confirmed as statistically significant using a paired t-test.