Scholarlink

<p>Biometric authentication systems, which use unique biological traits for identification, have gained popularity in various fields as a replacement for traditional password- or token-based systems. While offering enhanced security, these systems remain vulnerable to hacking attempts. To address this concern, this paper presents a novel masking technique for speaker identification, enhancing security and privacy. The core concept involves generating an alternate version of the original speaker template using a one-way transformation. This ensures that&#xa0;the original template remains protected and unused within the system. The transformation draws inspiration from deconvolution methods, commonly used for signal processing. These methods are applied in the presence of noise to magnify the noise in order to mask the original templates. We explore three specific deconvolution methods for creating cancelable biometric templates, namely Linear Minimum Mean Square Error (LMMSE), inverse filter and regularized deconvolution.&#xa0;The LMMSE method depends on statistical techniques to minimize the error between the original signal&#xa0;and the deconvolution output. The paper discusses some&#xa0;assumptions that help reduce the computational complexity of the LMMSE solution. Inverse filter deconvolution directly reverses the convolution process through multiplication with the inverse filter transfer function in frequency domain. Finally, regularized deconvolution incorporates additional constraints during the deconvolution process to improve the robustness of the transformed templates. Additionally, we investigate two further methods for comparison: Discrete Wavelet Transform (DWT) and wavelet thresholding. These methods offer alternative approaches to signal processing and feature extraction. Simulation results demonstrate the effectiveness of the deconvolution-based cancelable biometric schemes, displaying their potential for enhancing the security and privacy of speaker identification systems. A novel finding from this research is that the inverse filter deconvolution method, despite its known limitations, offers superior performance in comparison to other methods when applied in cancelable speaker identification. This unexpected outcomes change the conventional understanding of the method applicability and opens new directions for research in the field.</p>

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

Cancelable Speaker Identification Based on Speech Deconvolution Methods

  • 期刊论文

  • 发表日期：

    2025年11月27日

查看全文
  • Marwa A. Elsayed,
  • Walid El-Shafai,
  • Mohsen A. Rashwan,
  • Moawad I. Dessouky,
  • Adel S. El-Fishawy,
  • Fathi E. Abd El-Samie

摘要

Biometric authentication systems, which use unique biological traits for identification, have gained popularity in various fields as a replacement for traditional password- or token-based systems. While offering enhanced security, these systems remain vulnerable to hacking attempts. To address this concern, this paper presents a novel masking technique for speaker identification, enhancing security and privacy. The core concept involves generating an alternate version of the original speaker template using a one-way transformation. This ensures that the original template remains protected and unused within the system. The transformation draws inspiration from deconvolution methods, commonly used for signal processing. These methods are applied in the presence of noise to magnify the noise in order to mask the original templates. We explore three specific deconvolution methods for creating cancelable biometric templates, namely Linear Minimum Mean Square Error (LMMSE), inverse filter and regularized deconvolution. The LMMSE method depends on statistical techniques to minimize the error between the original signal and the deconvolution output. The paper discusses some assumptions that help reduce the computational complexity of the LMMSE solution. Inverse filter deconvolution directly reverses the convolution process through multiplication with the inverse filter transfer function in frequency domain. Finally, regularized deconvolution incorporates additional constraints during the deconvolution process to improve the robustness of the transformed templates. Additionally, we investigate two further methods for comparison: Discrete Wavelet Transform (DWT) and wavelet thresholding. These methods offer alternative approaches to signal processing and feature extraction. Simulation results demonstrate the effectiveness of the deconvolution-based cancelable biometric schemes, displaying their potential for enhancing the security and privacy of speaker identification systems. A novel finding from this research is that the inverse filter deconvolution method, despite its known limitations, offers superior performance in comparison to other methods when applied in cancelable speaker identification. This unexpected outcomes change the conventional understanding of the method applicability and opens new directions for research in the field.