Elaboration d'un système biométrique multimodale hybride
No Thumbnail Available
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Université Badji Mokhtar Annaba
Abstract
This thesis investigates the development of a multimodal biometric system that integrates electrocar diogram (ECG) and voice data to enhance the accuracy and reliability of person identification. The research aims to address limitations in unimodal biometric systems by combining complementary modalities, offering greater robustness against spoofing and improving overall performance. The study leverages the unique characteristics of both ECG signals and voice features, which are known to be resilient to environmental changes and health conditions, making them ideal for secure biometric authentication systems. The methodology involved two distinct unimodal systems for ECG and speaker recognition. The ECG-based system utilized a deep learning model, specifically GRU and LSTM architectures, trained on three major ECG databases (MIT-BIH, NSRDB, and PTB). The speaker recognition system employed a CNN-based model trained on a subset of 47 speakers from the LibriSpeech dataset. Both systems extracted meaningful features from the data (MFCC for voice and IMFs for ECG) after appropriate preprocessing steps. A multimodal biometric system was then developed by simulating a dataset combining voice and ECG samples, with score fusion techniques like Softmax and SVM used to merge the outputs of the unimodal systems. The key findings demonstrate that the GRU-based ECG system achieved an accuracy of 98.57% on the MIT-BIH database, outperforming the LSTM model. The CNN-based speaker recognition system reached an accuracy of 98.42%. However, the most significant result came from the multimodal system, where the fusion of ECG and voice scores using Softmax with the Sum rule yielded a high accuracy of 99.61%, with an EER of 0.22%. These results confirm that the multimodal approach provides superior performance compared to unimodal systems.
Description
Keywords
biometric identification; multimodal biometrics; ECG-based identification; speaker recognition; deep learning; GRU; LSTM; CNN; MFCC; empirical mode decomposition; score fusion; softmax; support vector machine