Deep learning based bio-metric authentication system using a high temporal/frequency resolution transform.

Area of Science:

• Biometrics
• Signal Processing
• Machine Learning

Background:

• Identity verification is critical in modern society, driving demand for robust authentication systems.
• Biometric modalities offer high accuracy and resistance to falsification, gaining significant attention.
• Electrocardiogram (ECG) signals possess unique, individualized characteristics suitable for biometric applications.

Purpose of the Study:

• To propose and evaluate a novel identity verification framework utilizing electrocardiogram (ECG) signals.
• To assess the framework's performance using benchmark datasets (NSRDB and MITDB) and deep learning techniques.

Main Methods:

• A two-step framework involving signal cleansing and frequency-domain transformation using the Wigner-Ville distribution.
• Conversion of ECG signals into image data for feature extraction, capturing unique cardiac signal information.
• Application of deep learning, specifically the GoogleNet architecture (a type of Convolutional Neural Network - CNN), for recognition.

Main Results:

• Achieved 99.3% accuracy and 0.8% Equal Error Rate (EER) on the NSRDB dataset.
• Demonstrated 99.004% accuracy and 0.8% EER on the MITDB dataset.
• Outperformed alternative biometric authentication methods in terms of accuracy and robustness.

Conclusions:

• ECG signals are effective for identity verification, offering high accuracy and noise resistance.
• The proposed framework, combining Wigner-Ville distribution and GoogleNet, shows the potential of deep learning in biometrics.
• The approach demonstrates high reliability and low error rates, with potential for extension and integration with other security layers.