Core Concepts
Existing audio deepfake detection models are vulnerable to simple manipulation attacks, such as volume control and fading, which can significantly bypass detection without affecting human perception. To address this, the proposed CLAD model leverages contrastive learning and length loss to enhance robustness against manipulation attacks.
Abstract
The paper presents a comprehensive study on the robustness of widely adopted audio deepfake detection models against various manipulation attacks. The authors find that even simple manipulations like volume control and fading can significantly bypass detection without affecting human perception.
To address this, the authors propose CLAD (Contrastive Learning-based Audio deepfake Detector), which incorporates contrastive learning to minimize the variations introduced by manipulations and enhance detection robustness. Additionally, CLAD employs length loss to improve the detection accuracy by clustering real audios more closely in the feature space.
The authors evaluate the performance of the most widely adopted audio deepfake detection models and CLAD against different manipulation attacks. The results show that the detection models are vulnerable, with the False Acceptance Rate (FAR) rising to 36.69%, 31.23%, and 51.28% under volume control, fading, and noise injection, respectively. In contrast, CLAD enhances robustness, reducing the FAR to 0.81% under noise injection and consistently maintaining an FAR below 1.63% across all tests.
Stats
The FAR of RawNet2 increases from 4.60% to 36.62% under volume control with a factor of 0.1.
The FAR of Res-TSSDNet increases from 1.63% to 51.28% under white noise with 15dB SNR.
The FAR of AASIST increases from 0.83% to 31.23% under fading with a half sinusoidal shape and 0.5 ratio.
Quotes
"Even manipulations like volume control can significantly bypass detection without affecting human perception."
"CLAD enhanced robustness, reducing the FAR to 0.81% under noise injection and consistently maintaining an FAR below 1.63% across all tests."