Scholarlink

The integration of liveness detection mechanisms in biometric systems is crucial for countering spoofing attacks, enhancing security, and optimizing lifestyle applications. This study explores the utilization of Photoplethysmography (PPG) signals for this purpose, leveraging their inherent liveness properties and cost-effectiveness. In the preprocessing stage, PPG signals of an online dataset is transformed into three different conversion methods, such as PPG signal representation, spectrograms, and Gram matrices. Five diverse models, ResNet50V2, DenseNet201, EfficientNetV2_B0, ConvNextBase, and an EfficientNetV2_B0 model with LSTM integration were training on the online dataset to be able to choose the best conversion technique into two dimensional format with the best model for biometric recognition. Remarkably high accuracies were achieved across multiple models on the online dataset, with the standout performer being the EfficientNetV2_B0 model integrated with LSTM, employing Gram matrix as a conversion technique which achieve 99% for the testing accuracy. This holistic approach has not only deepened our understanding of PPG signal processing but also highlighted the potential of deep learning models in future advancements in biometric identification and beyond.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Biometric Recognition with PPG Signals: A Comparative Study of Deep Learning Models and 2D Conversion Techniques

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ali Cherry,
  • Aya Nasser,
  • Mohamad Abou Ali,
  • Wassim Salameh,
  • Hadi Ballout,
  • Mohamad Hajj-Hassan

摘要

The integration of liveness detection mechanisms in biometric systems is crucial for countering spoofing attacks, enhancing security, and optimizing lifestyle applications. This study explores the utilization of Photoplethysmography (PPG) signals for this purpose, leveraging their inherent liveness properties and cost-effectiveness. In the preprocessing stage, PPG signals of an online dataset is transformed into three different conversion methods, such as PPG signal representation, spectrograms, and Gram matrices. Five diverse models, ResNet50V2, DenseNet201, EfficientNetV2_B0, ConvNextBase, and an EfficientNetV2_B0 model with LSTM integration were training on the online dataset to be able to choose the best conversion technique into two dimensional format with the best model for biometric recognition. Remarkably high accuracies were achieved across multiple models on the online dataset, with the standout performer being the EfficientNetV2_B0 model integrated with LSTM, employing Gram matrix as a conversion technique which achieve 99% for the testing accuracy. This holistic approach has not only deepened our understanding of PPG signal processing but also highlighted the potential of deep learning models in future advancements in biometric identification and beyond.