Scholarlink

<p>Speaker recognition systems (SRS) have become a central component of modern biometric systems, enabling non-intrusive user identification and authentication. However, despite their importance, these systems remain highly vulnerable to various forms of malicious attacks. These include replay attacks, speech synthesis, and, more recently, adversarial attacks, which subtly manipulate audio signals to deceive recognition models without being detected by the human ear. In this paper, we propose an innovative approach to strengthen SRS security. First, we introduce a ResNet-based attack model designed to generate targeted adversarial disruptions to speech recognition systems. Second, we develop a defense method based on audio source separation, allowing for the filtering and attenuation of disruptions before they are processed by the recognition system. To validate our approach, we conducted experiments on the Librispeech dataset, simulating different noise levels from the NOISEX-92 database to reproduce realistic acoustic conditions. The experimental results show that the proposed defense method significantly increases the robustness and scurity of SRS against three types of attacks FGSM, PGD and BIM.</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

Robust Speaker Verification System Against Adversarial Attacks Using ResNet and Source Separation in Noisy Conditions

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Yacine Abdelhakim Mokdad,
  • Ahmed Krobba,
  • Mohamed Debyeche

摘要

Speaker recognition systems (SRS) have become a central component of modern biometric systems, enabling non-intrusive user identification and authentication. However, despite their importance, these systems remain highly vulnerable to various forms of malicious attacks. These include replay attacks, speech synthesis, and, more recently, adversarial attacks, which subtly manipulate audio signals to deceive recognition models without being detected by the human ear. In this paper, we propose an innovative approach to strengthen SRS security. First, we introduce a ResNet-based attack model designed to generate targeted adversarial disruptions to speech recognition systems. Second, we develop a defense method based on audio source separation, allowing for the filtering and attenuation of disruptions before they are processed by the recognition system. To validate our approach, we conducted experiments on the Librispeech dataset, simulating different noise levels from the NOISEX-92 database to reproduce realistic acoustic conditions. The experimental results show that the proposed defense method significantly increases the robustness and scurity of SRS against three types of attacks FGSM, PGD and BIM.