mmHSE: Enhanced Eavesdropping Attack on Headsets Leveraging COTS mmWave Radar

Recent advancements in headset technology, coupled with the rising demand for secure, high-quality audio, have made headsets essential in both professional and personal contexts. However, emerging research highlights the vulnerability of these devices to sophisticated eavesdropping attacks using mmWave signals, where subtle diaphragm vibrations can unintentionally reveal playback speech content. Current eavesdropping methods face challenges, including low accuracy in modeling vibration signals, insufficient resolution in capturing diaphragm vibrations, and audio artifacts from nonlinear headset interactions, with FFT-based techniques further limited by resolution. To overcome these limitations, we propose the prototype system, namly mmHSE, which features two key innovations. First, we propose the Spatio-Temporal Segmented Fitting (STSF) method, which enhances signal clarity by isolating primary diaphragm vibrations through spatiotemporal signal segmentation. Second, we devise a two-stage integrated approach combining Minimum Variance Distortionless Response (MVDR) beamforming with the Chirp-Z Transform (CZT), achieving a resolution of 0.04 cm and significantly enhancing vibration capture precision. Additionally, mmHSE incorporates a conditional Generative Adversarial Network (cGAN) to refine Mel Spectrograms, enabling effective speech recovery. Experimental results demonstrate that mmHSE outperforms existing methods, with a 15.4% increase in Peak Signal-to-Noise Ratio (PSNR), an 11.4% reduction in Mel Cepstral Distortion (MCD), a 14.5% improvement in Word Error Rate (WER), and approximately a 15.1% increase in Likert Score, confirming its effectiveness in enhancing eavesdropping accuracy on headsets.