debbis
Deep Electrocardiogram Based Biometric Identification System
Axel Mukwena, DB727501, Group 20 | Supervised by Prof. Liming Zhang
2021 FYP, Computer Science, Faculty of Science and Technology
Overview
● Objectives
● Background
● System Structure
● Data Processing
● Neural Network Architectures
● Results
● Challenges
● Conclusions
Objectives
● Explore ECG biometrics using Deep Learning: CNN
and Siamese NN.
Why ECGs?
● Universal, unique and guarantees liveness
detection
● Harder to replicate*
● Passwords, pins etc. can get lost, stolen or can be
guessed
● Fingerprints [1], iris [2] and facial features [3] can
be falsified or reverse engineered from templates
Figure 1: Steps during fingerprint reverse engineering from ISO
templates. Source [1]
Background
What is Biometric Recognition (BR)?
● Establishing an identity based on physiological or
physical data.
● Current works use [4]:
○ Fiducial method i.e P, Q, R, S and T
○ Non-fiducial methods
○ Partially fiducial methods i.e R-peak detection
Figure 2: Representation of the QRS complex [Wikipedia]
● Existing Deep Learning implementations [5]
○ Generic and Temporal CNN
○ RNN
○ Autoencoders
System Structure
Stages of Biometric System
● Enrolment
○ Saving user details on database
○ Creating templates
● Recognition
○ Verification and or Authentication
○ Compare test samples to templates and
output a match
System Structure
Figure 3: a) System flow structure with verification and authentication
System Structure
Figure 3: b) System flow structure with verification
Data processing
Data Sources
● MIT-BIH Physionet
Figure 4: a) Shows the BMD device used to extract signals; b)
● ECG-ID Physionet Demonstrate live signal Acquisition
● BMD101 Sensor Chip
○ Non-intrusive
○ Subject to hold the two dry
metal electrodes with each
thumb
Figure 5: BMD Windows-OS Graphical User Interface
Data processing
ECG Segmentation
● No filters applied
Figure 6: Unsegmented ECG wave section
● R-peak segmentation using
Christov algorithm [6]
● Resampled to 256
Figure 7: One ECG segment of size 256
Data processing
Data Augmentation
● Gaussian Noise
● Time Shifting
● Pitch Shifting
Figure 8: One ECG segment of size 256
Neural Network Architectures
Block
Convolution Neural Network
● 4 block structure inherited from
VGG19 [7] architecture
● 256-size 1D signal input
Figure 9: Block structure used in
the CNN
Figure 10: Convolution Neural Network
Architecture
Neural Network Architectures
Block
Siamese Neural Network
● 5 block structure also inherited
from VGG19 [7] architecture
● 256-size 1D signal pair input
○ Positive and Negative pairs
Uses Euclidean Distance for
Figure 11: Block structure used in
● SNN
similarity score
Figure 12: Siamese Neural Network Architecture
Results
SNN CNN
● Data from MIT-BIH and two Input Type Segment Pair 1 Segment
classes from BMD101 Sensor Sample size 400 000 397 740
● ECG-ID dataset failed to produce Training time 13 min 54 min
materialistic results due to Valid Acc - 0.9993
limited data.
Valid Loss 0.4969 0.0029
Threshold, 0.000999 0.99
Decision Margin
Figure 10: Convolution Neural Network Architecture
Results
● SNN Training
Figure 13: a) Training results for the SNN. In Blue is the Training loss
Figure 13: b) Predicted test samples from the SNN model
and Red is Validation Loss:
Results
● CNN Training
Figure 14: a) Training results for the CNN. In Blue is the Training loss
Figure 14: b) Predicted test samples from the CNN model
and Red is Validation Loss:
Challenges
● Closed-set vs Open-set classification
● Binary Classification vs Multi-class
classification
● Subjects with cardiac diseases have non-
uniform ECGs
● ECGs are susceptible to noise
Conclusions
● Increased efforts for better authentication
system and overcoming their limitations.
● 3D face recognition used to counteract 2D
limitations, but 3D models are also
vulnerable to attacks [8].
References
[1] Cappelli et al (2006, December). Can Fingerprints be reconstructed from
ISO Templates?.
[2] Venugopalan et al (2013). Iris spoofing: Reverse engineering the daugman
feature encoding scheme.
[3] Mai et al. (2018). On the reconstruction of face images from deep face
templates.
[4] Hassan et al. Review of Fiducial and Non-Fiducial Techniques of Feature
Extraction in ECG based Biometric Systems
[5] Eduardo et al. ECG-based biometrics using a deep autoencoder for feature
learning-an empirical study on transferability.
[6] Christov, I. I. (2004). Real time electrocardiogram QRS detection using
combined adaptive threshold.
[7] Simonyan et al (2014). Very deep convolutional networks for large-scale
image recognition.
[8] Jia et al (2019). A database for face presentation attack using wax figure
faces.
The End Q & As
