BlowLive
Blow-based multi-factor biometrics with liveness detection and revocability (ESORICS 2026). 99.46% liveness-detection accuracy on live versus replayed blows.
- Android
- PyTorch
- CNN
- Triplet loss
- FaceNet
- Fuzzy extractor
- Doppler analysis
- Ultrasound
Extends the blow-and-face approach with liveness detection and revocability. A 1D CNN trained with triplet loss turns gammatone-based audio features of the blow into a 128-dimensional embedding, and a fine-tuned FaceNet model provides the face embedding. A fuzzy extractor with error-correcting codes derives a stable cryptographic key from the noisy biometric, so the key can be revoked by issuing new helper data. For liveness, the Android app plays a near-inaudible 20 kHz tone while the user blows and records the reflection; the Doppler shift caused by the airflow is used to tell a live blow from a replayed one. Data was collected from a new group of 50 participants, separate from the BlowPrint group, including the ultrasound recordings used for liveness. Reported results: 99.56% authentication accuracy for blow acoustics, 100% for the face and for fusion, and 99.46% liveness-detection accuracy. Paper: "BlowLive: Blow-Based Multi-Factor Biometrics with Liveness Detection and Revocability" (ESORICS 2026).