3D-thief: exploiting side-channel vulnerability in modern 3D printers

• Christos Madamopulos, Electrical Engineering, National Technical University of Athens

• Nektarios Tsoutsos, Electrical and Computer Engineering, University of Delaware

Securing the cyberspace has evolved as one of the most complex engineering challenges we are facing today. In particular, Side Channel Attacks (SCA) are an increasing threat to different cyber-physical systems. Side channels are methods used to gather information about a system or its operations without directly interacting with the system’s main functionality. These attacks are based on side channel data (e.g., sound), which can be correlated to the system’s main functionality, and hence compromise its security.
The project focuses on the evaluation of side channel data, i.e., sound, of modern 3D printers for vulnerabilities. The project also aims to create a testbench to allow further research in the area of cyber-physical system security and explore the vulnerability of fused deposition modeling to SCAs.
In particular, time and frequency domain characteristics were extracted from the audio recording of a 3D printer. These characteristics are used by a system of machine learning models to detect part of the printing head’s motion, through regression models for the prediction of the speed, in different axes, and classification models for change in speed and change in direction. To better be able to evaluate the machine learning models to find the best one for each application, a testbench was developed. The user can input Gcode alongside recordings to synthesize audio signals modeled after those recordings (speeds, change in direction, change in speed). The synthesized signals are then supplied to a user chosen machine learning model, to predict specific parts of the simulated nozzle motion. The results are used to generate Gcode, and compare it to the original Gcode, getting the effectiveness of the machine learning model on the specific task.