A Non-Line-of-Sight, Multi-Modality-based Side-Channel IP Theft Attack on Additive Manufacturing Using Dual Smartphones

Technical Summary: Non-Line-of-Sight, Multi-Modality-based Side-Channel IP Theft Attack on Additive Manufacturing Using Dual Smartphones

Problem Statement
Additive Manufacturing (AM), or 3D printing, has revolutionized sectors such as aerospace, healthcare, and automotive by enabling rapid prototyping and on-demand production. However, this widespread adoption introduces significant Intellectual Property (IP) risks. The printing process generates unintended side-channel emissions—specifically acoustic and magnetic signals—that correlate with the underlying G-code instructions. While prior research has demonstrated the feasibility of reconstructing G-code from these emissions, existing studies face three critical limitations:

1. Low Success Rates: Previous attempts often struggle with high accuracy in reconstructing final G-code commands.
2. Limited Range: Attacks typically require close-range placement (less than 30 cm) and direct line-of-sight (LOS).
3. Specialized Equipment: Many approaches rely on overt, specialized data-collection tools rather than commodity devices.
   Furthermore, prior evaluations often assume simplified operational behaviors (ignoring idle cycles or cooling) and controlled laboratory environments with minimal background noise, which do not reflect realistic manufacturing settings where occlusion, concurrent machine activity, and ambient noise are prevalent.

Methodology
This paper proposes a practical, non-invasive side-channel attack utilizing two commodity smartphones to reconstruct G-code from a 3D printer without physical contact or line-of-sight.

• Threat Model: The attacker possesses general knowledge of the printer model (Fused Deposition Modeling) but has no access to firmware, internal telemetry, or original design files. The attacker deploys two hidden smartphones equipped with microphones and magnetometers at a non-LOS distance of 60 cm from the printer's surface bed, positioned at a 45° angle.
• Data Acquisition: The attack captures two distinct signal modalities:
  • Acoustic: Generated by mechanical motion and extrusion activity.
  • Magnetic: Generated by stepper motor drivers and moving components.
    The 45° angle was selected as a compromise to maintain sufficient acoustic amplitude (following a cosine projection model) while maximizing magnetic observability through axis mixing.
• Signal Processing: Since the two smartphones operate independently, the raw data streams are unsynchronized. The methodology employs a synchronization pipeline that aligns timestamps, interpolates missing values, and normalizes the data. The combined input vector consists of eight dimensions (three magnetic axes and one acoustic value from each of the two phones).
• Machine Learning Model: The preprocessed data is segmented into overlapping time windows and fed into a hybrid CNN-LSTM architecture.
  • CNN Layers: Extract spatial and spectral features from the emissions.
  • LSTM Layer: Captures temporal dependencies and the sequence of printer states (e.g., moving, idle, extruding).
  • Output: The model classifies motion primitives (IDLE, UP, DOWN, LEFT, RIGHT, Z_UP, Z_DOWN) and extrusion indicators, which are then aggregated to reconstruct the G-code command stream.

Key Contributions

1. Dual-Smartphone Multi-Modal Attack: The study demonstrates a concealed attack using two smartphones and a dynamic noise modeling framework, enabling G-code reconstruction in noisy environments, including scenarios where a second 3D printer operates concurrently.
2. Non-Line-of-Sight Reconstruction: The attack succeeds without physical contact or LOS, utilizing placements 60 cm away from the printer.
3. High-Fidelity Reconstruction: The method achieves a command-level reconstruction accuracy of 98.89%, producing a printed object that geometrically matches the original. The approach maintains high accuracy at larger distances compared to prior works, reducing dimensional and surface-quality errors.
4. Robustness Validation: The methodology was tested in a different physical environment with a different printer unit of the same brand, successfully reconstructing complex objects (gear icon, turbine blades, wrench) with accuracies ranging from 98.05% to 98.35%.
5. Mitigation Guidance: The paper outlines practical countermeasures, including acoustic/magnetic shielding (soundproof enclosures, Faraday cages), randomization of printing parameters (path obfuscation, speed variation), and active jamming.

Results

• Primary Experiment: In the main scenario (60 cm distance, non-LOS, concurrent noise from a second printer), the attack achieved 98.89% command-level accuracy. The reconstructed G-code preserved the original geometry with only minor discrepancies in feed rates and extrusion values.
• Interference Analysis: When a second printer (Creality) operated simultaneously with the target (LulzBot), the target's signals remained dominant due to larger motors and louder acoustic emissions. The magnetic signals from the target were also stronger and more consistent due to specific motor placement relative to the sensors.
• Distance Sensitivity: Accuracy remained high (98–99%) at distances up to 70 cm. Performance degraded significantly beyond 100 cm as the Signal-to-Noise Ratio (SNR) dropped, with accuracy falling below 65% at distances greater than 150 cm.
• Model Performance: The hybrid CNN-LSTM model achieved a training accuracy of 97.99% on motion classification, outperforming SVM, Random Forest, standalone CNN, and standalone LSTM models.

Significance
The paper claims to demonstrate that 3D printing systems possess significant security weaknesses regarding side-channel leakage. By proving that commodity smartphones can be used to unauthorizedly access and reconstruct high-fidelity G-code from a distance without physical contact, the study highlights the vulnerability of AM processes in shared or public environments. The authors emphasize that the ease of data collection using ubiquitous devices necessitates immediate awareness and the implementation of mitigating countermeasures to protect manufacturing IP.