Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WeDefense: A Toolkit to Defend Against Fake Audio

Lin Zhang, Johan Rohdin, Xin Wang, Junyi Peng, Tianchi Liu, You Zhang, Hieu-Thi Luong, Shuai Wang, Chengdong Liang, Anna Silnova, Nicholas Evans

TL;DR

WeDefense addresses the rising threat of fake audio by providing the first open-source toolkit to jointly defend against detection and localization of synthetic audio. The framework unifies datasets, metrics, input modalities, augmentation, and model types (including SSL-based backends) under a modular, deployment-ready platform, enabling fair benchmarking and reproducible research. Key contributions include calibrated score fusion, comprehensive analyses (UMAP, Grad-CAM), and support for both partial and full spoofing scenarios across PartiaIspoof and ASVspoof5 datasets. The toolkit lowers barriers to entry, facilitates cross-model comparisons, and supports practical deployment, with interactive demos and TorchScript export to bridge research and real-world use. Overall, WeDefense advances reproducible, extensible, and transparent evaluation of fake audio defenses in a rapidly evolving generative-AI landscape.

Abstract

The advances in generative AI have enabled the creation of synthetic audio which is perceptually indistinguishable from real, genuine audio. Although this stellar progress enables many positive applications, it also raises risks of misuse, such as for impersonation, disinformation and fraud. Despite a growing number of open-source fake audio detection codes released through numerous challenges and initiatives, most are tailored to specific competitions, datasets or models. A standardized and unified toolkit that supports the fair benchmarking and comparison of competing solutions with not just common databases, protocols, metrics, but also a shared codebase, is missing. To address this, we propose WeDefense, the first open-source toolkit to support both fake audio detection and localization. Beyond model training, WeDefense emphasizes critical yet often overlooked components: flexible input and augmentation, calibration, score fusion, standardized evaluation metrics, and analysis tools for deeper understanding and interpretation. The toolkit is publicly available at https://github.com/zlin0/wedefense with interactive demos for fake audio detection and localization.

WeDefense: A Toolkit to Defend Against Fake Audio

TL;DR

WeDefense addresses the rising threat of fake audio by providing the first open-source toolkit to jointly defend against detection and localization of synthetic audio. The framework unifies datasets, metrics, input modalities, augmentation, and model types (including SSL-based backends) under a modular, deployment-ready platform, enabling fair benchmarking and reproducible research. Key contributions include calibrated score fusion, comprehensive analyses (UMAP, Grad-CAM), and support for both partial and full spoofing scenarios across PartiaIspoof and ASVspoof5 datasets. The toolkit lowers barriers to entry, facilitates cross-model comparisons, and supports practical deployment, with interactive demos and TorchScript export to bridge research and real-world use. Overall, WeDefense advances reproducible, extensible, and transparent evaluation of fake audio defenses in a rapidly evolving generative-AI landscape.

Abstract

The advances in generative AI have enabled the creation of synthetic audio which is perceptually indistinguishable from real, genuine audio. Although this stellar progress enables many positive applications, it also raises risks of misuse, such as for impersonation, disinformation and fraud. Despite a growing number of open-source fake audio detection codes released through numerous challenges and initiatives, most are tailored to specific competitions, datasets or models. A standardized and unified toolkit that supports the fair benchmarking and comparison of competing solutions with not just common databases, protocols, metrics, but also a shared codebase, is missing. To address this, we propose WeDefense, the first open-source toolkit to support both fake audio detection and localization. Beyond model training, WeDefense emphasizes critical yet often overlooked components: flexible input and augmentation, calibration, score fusion, standardized evaluation metrics, and analysis tools for deeper understanding and interpretation. The toolkit is publicly available at https://github.com/zlin0/wedefense with interactive demos for fake audio detection and localization.
Paper Structure (29 sections, 4 figures, 7 tables)

This paper contains 29 sections, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Fake Audio Detection and Localization
  3. Fake Audio Detection
  4. Fake Audio Localization
  5. WeDefense
  6. Overall Structure
  7. Databases - focus on both partially and fully spoofed audio
  8. Input and Augmentation - flexible for defending against fake audio
  9. Models - support both conventional models and SSL-based models, detection and localization tasks.
  10. Loss Functions - including losses specific for defending against fake audio
  11. Calibration and Fusion
  12. Metrics
  13. Detection
  14. Localization
  15. Analyses
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Overall Structure of WeDefense.
  • Figure 2: Interaction Demo of Fake Audio Localization in WeDefense.
  • Figure 3: Example of Embedding Visualization by UMAP on the dev. set of ASVspoof5 (annotated according to the released metadata).
  • Figure 4: Frame-level Grad-CAM visualization of parts of a segment from CON_E_0033629.wav in the PartialSpoof evaluation set. Regions with darker red indicate higher model contribution.