Asset Leakage with Acoustic Side-Channel Attacks in Multi-Joint Collaborative Robots.
Collaborative robots (cobots) are increasingly deployed in high-value manufacturing environments, where their motion trajectories encode proprietary intellectual property (e.g., aerospace composite layups, pharmaceutical compounding). While prior acoustic attacks targeted stepper motors in 3D printers, ALASCA presents the first acoustic side-channel attack targeting the continuous, broadband noise of brushless DC (BLDC) servo motors and harmonic drives in modern cobots.
Using a single four-channel Ambisonic microphone (Zoom H3-VR) positioned near the robot, we capture omnidirectional and directional particle velocity (X, Y, Z axes). By analyzing how acoustic energy radiates asymmetrically during rotation, we can extract spatial features that provide a critical 14.4% accuracy boost over traditional single-channel recording.
We extract a 270-dimensional feature vector per 250ms window, capturing spectral (MFCCs, Chroma), spatial (energy ratios), and temporal (Delta MFCCs) dynamics. These features train Random Forest classifiers that generalize across vastly different motion profiles, learning the underlying physics of motor commutation and mechanical load.
The attack successfully extracts multi-joint motion parameters purely from sound. We achieve 78.4% accuracy for rotation direction, 94.8% accuracy for angular velocity (±1 RPM), and 95.9% accuracy in identifying which joint is moving. Astonishingly, the classifier can also identify the underlying control algorithm (Bang-Bang, PID, or Predictive) with 99.99% accuracy, enabling end-to-end proprietary trajectory reconstruction (13.1° RMSE).