NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation

Huhong Xian, Rui Liu, Berrak Şişman, Haizhou Li

Abstract

Different from traditional sentence-level audio deepfake detection (ADD), partial audio deepfake detection (PADD) requires frame-level positioning of the location of fake speech. While some progress has been made in this area, leveraging semantic information from audio, especially named entities, remains an underexplored aspect. To this end, we propose a novel method, NE-PADD, which leverages named entity knowledge for robust partial audio deepfake detection through attention aggregation. NE-PADD consists of two parallel branches: Speech Named Entity Recognition (SpeechNER) and PADD. Specifically, we introduce two attention aggregation mechanisms to help PADD models better understand named entity knowledge, thus achieving more robust PADD performance. (1) Attention Fusion (AF) combines attention from SpeechNER and PADD for more accurate weights. (2) Attention Transfer (AT) uses an auxiliary loss from their attention distributions to guide PADD in learning named entity semantics. We construct an appropriative PartialSpoofNER dataset based on the existing PartialSpoof dataset and conduct a detailed comparative analysis of the two attention aggregation methods. Experiments demonstrate that our method exceeds all advanced baselines and demonstrates the effectiveness of fusing named entity knowledge during PADD. The code is available at https://github.com/AI-S2-Lab/NE-PADD.
BibTeX
@misc{xian2025ne,
  author = {Huhong Xian and Rui Liu and Berrak Şişman and Haizhou Li},
  title = {{NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation}},
  year = {2025},
  url = {https://openreview.net/forum?id=tVg3d5pTj9},
}

Loading related content...

Figure