Professor, California Polytechnic State University
Shipped: Side-Channel Attacks Survive Noise Cancellation in 3D Printers
Shows that ML models can still recover 3D printer activity from acoustic and vibration leakage, even with noise cancellation. Published in IEEE Access, 2026.
side-channelhardware-securitymachine-learningcyber-physical-systemsadditive-manufacturing






Comments2
Noise cancellation mostly removes the dominant stationary components, but the stepper and extruder harmonics probably still carry the signal. Did the models lean on specific frequency bands, and how well did they transfer across printer models or microphone placement?
Great question, Ryan. Our spectral analysis identified residual leakage in the 120–360 Hz motor-frequency band despite active noise cancellation. One interesting finding was that print duration alone achieved 63.89% accuracy in identifying 12 printed-object classes, compared with an 8.33% chance baseline. This was a timing side channel, independent of acoustic or vibration features. We also found that cross-printer transfer between the Bambu P1P and A1 Mini remained near chance, suggesting architecture-specific leakage. Microphone placement was not systematically evaluated and remains an area for future research.
2 comments
Noise cancellation mostly removes the dominant stationary components, but the stepper and extruder harmonics probably still carry the signal. Did the models lean on specific frequency bands, and how well did they transfer across printer models or microphone placement?
Great question, Ryan. Our spectral analysis identified residual leakage in the 120–360 Hz motor-frequency band despite active noise cancellation. One interesting finding was that print duration alone achieved 63.89% accuracy in identifying 12 printed-object classes, compared with an 8.33% chance baseline. This was a timing side channel, independent of acoustic or vibration features. We also found that cross-printer transfer between the Bambu P1P and A1 Mini remained near chance, suggesting architecture-specific leakage. Microphone placement was not systematically evaluated and remains an area for future research.