Scholarlink

<p>With the rapid advancement of artificial intelligence technologies, deepfake speech has emerged as a significant threat across domains including finance, politics, and public security. Synthetic audio content can propagate misinformation, cause financial losses, damage reputations, and even destabilize social order. Therefore, effective detection of deepfake speech is crucial for safeguarding information authenticity and maintaining social trust. However, existing Deepfake Speech Detection (DSD) algorithms often fail to meet the required robustness and inference efficiency when confronted with increasingly sophisticated generative attacks. To address these challenges, this paper proposes a more resilient and faster DSD framework. First, we design a feature extraction front-end based on an encoder-only Transformer architecture, enabling the learning of more robust speech representations. Second, we introduce a knowledge distillation framework built upon an Audio Transformer, comprising a teacher network, a monitorial network, and multiple student networks. By employing an online one-teacher–multi-students distillation strategy, our method not only enhances detection accuracy but also significantly accelerates inference speed, supporting real-time deployment. To improve resilience against future unknown and more complex spoofing attacks, we further train the model on the ASVspoof2021 LA dataset using pseudo-labels generated by the teacher model, while allowing the teacher to update its parameters online based on student feedback. Experimental results show that on the ASVspoof2019 evaluation set, our approach achieves state-of-the-art performance with the smallest model size. Compared with existing SOTA systems, our method improves accuracy by 8.8% while reducing the number of parameters by 24%. Moreover, on both the ASVspoof2021 LA and ASVspoof2021 DF evaluations, our model contains only 0.22&#xa0;M parameters and achieves superior performance over previously reported methods. Additionally, inference tests demonstrate that detection for a single utterance on a CPU takes only 7.64&#xa0;ms, satisfying real-time processing requirements.</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

The Online Knowledge Distillation Algorithm Based on Monitorial mechanism for Deepfake Speech Detection

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Weidong An,
  • Ruwei Li,
  • Huaiyu Li,
  • Haoyu Ge

摘要

With the rapid advancement of artificial intelligence technologies, deepfake speech has emerged as a significant threat across domains including finance, politics, and public security. Synthetic audio content can propagate misinformation, cause financial losses, damage reputations, and even destabilize social order. Therefore, effective detection of deepfake speech is crucial for safeguarding information authenticity and maintaining social trust. However, existing Deepfake Speech Detection (DSD) algorithms often fail to meet the required robustness and inference efficiency when confronted with increasingly sophisticated generative attacks. To address these challenges, this paper proposes a more resilient and faster DSD framework. First, we design a feature extraction front-end based on an encoder-only Transformer architecture, enabling the learning of more robust speech representations. Second, we introduce a knowledge distillation framework built upon an Audio Transformer, comprising a teacher network, a monitorial network, and multiple student networks. By employing an online one-teacher–multi-students distillation strategy, our method not only enhances detection accuracy but also significantly accelerates inference speed, supporting real-time deployment. To improve resilience against future unknown and more complex spoofing attacks, we further train the model on the ASVspoof2021 LA dataset using pseudo-labels generated by the teacher model, while allowing the teacher to update its parameters online based on student feedback. Experimental results show that on the ASVspoof2019 evaluation set, our approach achieves state-of-the-art performance with the smallest model size. Compared with existing SOTA systems, our method improves accuracy by 8.8% while reducing the number of parameters by 24%. Moreover, on both the ASVspoof2021 LA and ASVspoof2021 DF evaluations, our model contains only 0.22 M parameters and achieves superior performance over previously reported methods. Additionally, inference tests demonstrate that detection for a single utterance on a CPU takes only 7.64 ms, satisfying real-time processing requirements.