Handwritten signature verification using a wearable surface-EMG armband

Area of Science:

• Biometrics and Human-Computer Interaction
• Machine Learning and Signal Processing

Background:

• Remote authentication requires reliable signature verification.
• Handwritten signatures present significant intra-class variability, challenging traditional systems.
• Surface electromyography (sEMG) offers a potential biometric modality through wearable devices.

Purpose of the Study:

• To investigate the efficacy of sEMG signals captured by wearable armbands for signature verification.
• To develop and evaluate a dual-model deep learning framework for enhanced signature verification.
• To address the challenge of intra-class variability in handwritten signatures using sEMG data.

Main Methods:

• Collected 4-channel sEMG data from 20 individuals signing Chinese characters using a wearable armband.
• Developed a dual-model deep learning framework integrating muscle co-activation patterns and raw sEMG waveforms.
• The framework includes a CNN-LSTM for muscle activation sequences and a multi-branch CNN for raw sEMG signals, with decision-level fusion.

Main Results:

• Conventional feature-based methods achieved 80.90% accuracy and 12.82% equal error rate (EER).
• The proposed dual-model deep learning framework achieved 91.65% accuracy and 5.25% EER.
• Demonstrated a 10.75% improvement in accuracy and significant reduction in EER compared to traditional techniques.

Conclusions:

• The proposed sEMG-based dual-model deep learning framework effectively reduces intra-class variability in signature verification.
• This approach offers a practical and secure biometric authentication solution leveraging wearable technology.
• Findings highlight the potential of sEMG for robust and usable remote authentication systems.