“Assuring Security of STL Files for Additive Manufacturing”

Additive manufacturing is a multi-billion dollar industry (20.3 billion in 2023). This industry and its processes should be dependable and secure. In the case of security, malicious actors can inject or modify geometry in designs with STL format which describe a triangulated 3D object. These modifications can have a devastating impact on a final product’s quality. Current ways of detecting these modifications use machine sensor or simulation data, as well as physical verification measures after printing. However, to the best of our knowledge, no method exists for detecting these changes solely at the STL level. By leveraging the information provided by the STL file, we can design efficient and economic detection measures which have the potential to detect modifications in designs en masse. We can also leverage the STL format to uncover hidden information based on graph-theoretic properties, and assure the security of STL files based on such properties.

In this dissertation, we propose AVOID, a new approach to detect negative space in an STL file called hidden voids, and VSHIFT, a new approach to detect when points or vertices in the file have been subtly shifted.

AVOID both detects voids and assesses their risk of weakening the manufactured part. We empirically evaluated AVOID using several large datasets, in total thousands of STL files, each with multiple hidden voids. We found that AVOID is highly accurate, with 97.7% recall, 98.1% precision, and 99% F1 on average across six data sets. AVOID is also robust, scalable, and efficient. We find that high risk voids account for approximately 7% of all detected voids inserted randomly. For vertex shifting, we provide the first known formal definition of such an attack, and develop VSHIFT to detect shifted points in the file. We evaluated VSHIFT using a large data set of hundreds of STL files, similar to AVOID, each with at least one vertex shifted. We compared several machine learning models with our method and found that a stacking model with random forest and logistic regression performed best overall, with 93.6% precision, 90.6% recall, and 92.1% F1. Regarding STL graph properties, we define two classes of graphs, mesh graphs and facet graphs, and investigate several properties, such as bipartiteness and watertightness of the mesh, which can be leveraged in STL software to assure the status of designs. We prove several important theorems about facet graph connectivity and planarity, and we address how their presence may be leveraged to affect security measures in additive manufacturing.