Continual Audio Deepfake Detection via Universal Adversarial Perturbation

Authors: Wangjie Li, Lin Li, Qingyang Hong

Published: 2025-11-25 06:41:11+00:00

AI Summary

This paper introduces a novel framework for continual audio deepfake detection that leverages Universal Adversarial Perturbation (UAP). This approach allows models to retain knowledge of historical spoofing distributions without needing direct access to past data, addressing the challenge of evolving deepfake attacks and high fine-tuning costs. By integrating UAP with pre-trained self-supervised audio models, the method offers an efficient solution for continual learning.

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.

Key findings

The proposed UAP-based continual learning framework effectively mitigates catastrophic forgetting in audio deepfake detection models, achieving significant performance retention on historical datasets compared to standard sequence fine-tuning (up to 48% relative improvement). Experiments also demonstrated that feature-level UAP is more effective than waveform-level UAP in preserving prior knowledge and enhancing continual learning stability, due to its ability to capture less redundant details.

Approach

The authors propose a continual learning framework for audio deepfake detection that utilizes Universal Adversarial Perturbation (UAP). UAP, generated from historical models, is combined with bona fide samples to create pseudo-spoofed data, effectively preserving past spoofing distributions without storing old datasets. This UAP is integrated during the fine-tuning of pre-trained self-supervised audio models, further enhanced by knowledge distillation to maintain distributional consistency.

Datasets

ASVspoof 2019 LA, CFAD, ASVspoof 5, ASVspoof 2021 LA, ASVspoof 2021 DeepFake

Model(s)

WavLM (pre-trained self-supervised audio model based on CNN and Transformer encoders), followed by Transformer layers and a fully connected layer for classification.

Author countries

China

← Previous