Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continual Audio Deepfake Detection via Universal Adversarial Perturbation

Wangjie Li, Lin Li, Qingyang Hong

TL;DR

This work tackles the problem of continual audio deepfake detection under evolving spoofing attacks by introducing Universal Adversarial Perturbation (UAP) into a continual learning framework. The method generates a UAP from the previous model and uses it as a pseudo-spoofed signal during fine-tuning of a pre-trained self-supervised audio model (WavLM), with distillation to preserve prior distributions on real samples. The authors show that feature-level UAPs outperform waveform-level perturbations in maintaining memory of past spoofing distributions, achieving better cross-domain generalization across multiple public datasets (e.g., ASVspoof and CFAD) while avoiding storing historical data. Overall, the approach provides an efficient, data-leakage-free pathway to robust continual learning in audio anti-spoofing, with significant gains over standard sequence fine-tuning. The work also highlights the practical impact of adversarial perturbations as a discriminative feature for spoofing in the context of evolving threats.

Abstract

The rapid advancement of speech synthesis and voice conversion technologies has raised significant security concerns in multimedia forensics. Although current detection models demonstrate impressive performance, they struggle to maintain effectiveness against constantly evolving deepfake attacks. Additionally, continually fine-tuning these models using historical training data incurs substantial computational and storage costs. To address these limitations, we propose a novel framework that incorporates Universal Adversarial Perturbation (UAP) into audio deepfake detection, enabling models to retain knowledge of historical spoofing distribution without direct access to past data. Our method integrates UAP seamlessly with pre-trained self-supervised audio models during fine-tuning. Extensive experiments validate the effectiveness of our approach, showcasing its potential as an efficient solution for continual learning in audio deepfake detection.

Continual Audio Deepfake Detection via Universal Adversarial Perturbation

TL;DR

This work tackles the problem of continual audio deepfake detection under evolving spoofing attacks by introducing Universal Adversarial Perturbation (UAP) into a continual learning framework. The method generates a UAP from the previous model and uses it as a pseudo-spoofed signal during fine-tuning of a pre-trained self-supervised audio model (WavLM), with distillation to preserve prior distributions on real samples. The authors show that feature-level UAPs outperform waveform-level perturbations in maintaining memory of past spoofing distributions, achieving better cross-domain generalization across multiple public datasets (e.g., ASVspoof and CFAD) while avoiding storing historical data. Overall, the approach provides an efficient, data-leakage-free pathway to robust continual learning in audio anti-spoofing, with significant gains over standard sequence fine-tuning. The work also highlights the practical impact of adversarial perturbations as a discriminative feature for spoofing in the context of evolving threats.

Abstract

The rapid advancement of speech synthesis and voice conversion technologies has raised significant security concerns in multimedia forensics. Although current detection models demonstrate impressive performance, they struggle to maintain effectiveness against constantly evolving deepfake attacks. Additionally, continually fine-tuning these models using historical training data incurs substantial computational and storage costs. To address these limitations, we propose a novel framework that incorporates Universal Adversarial Perturbation (UAP) into audio deepfake detection, enabling models to retain knowledge of historical spoofing distribution without direct access to past data. Our method integrates UAP seamlessly with pre-trained self-supervised audio models during fine-tuning. Extensive experiments validate the effectiveness of our approach, showcasing its potential as an efficient solution for continual learning in audio deepfake detection.

Paper Structure

This paper contains 15 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Base training mechanism
  4. UAP training stage
  5. Model fine-tuning stage
  6. Loss function
  7. Experimental setup
  8. Datasets and metrics
  9. Model architecture
  10. Implementation details
  11. Results and analysis
  12. Baseline performance
  13. Sequence fine-tuning with UAP
  14. Impact of different UAP forms
  15. Conclusions

Figures (2)

  • Figure 1: Overview of proposed training framework for continual audio deepfake detection. Subfigure (a) details the UAP training stage and subfigure (b) elucidates the model fine-tuning stage.
  • Figure 2: Visualization of embedding distribution on ASVspoof 5 evaluation set, plotted using UMAP Umap dimension reduction. Left: feature-level UAP. Right: wavform-level UAP.