Authors:
Haolin Wu, Jing Chen, Ruiying Du, Cong Wu, Kun He, Xingcan Shang, Hao Ren, and Guowen Xu.
Where published:
Knowledge-Based Systems, Volume 335, Article 115179, 2026.
Abstract:
This paper studies how audio deepfake detection systems can be fooled by simple audio manipulations (like volume changes, fading, and noise), showing that even minor, human-imperceptible modifications can significantly degrade detection performance. To address this vulnerability, the authors propose a new method called CLAD (Contrastive Learning-based Audio Deepfake Detector). Their approach uses contrastive learning to make the model invariant to such manipulations by learning more stable and robust representations, and introduces an additional length loss to better cluster real (bonafide) audio in the feature space. Through extensive experiments, they demonstrate that existing detectors are highly vulnerable (with high false acceptance rates), while CLAD significantly improves robustness, reducing error rates across various attack scenarios. Overall, the paper both reveals critical weaknesses in current systems and proposes a more resilient detection framework.
Dataset names (used for):
- ASVspoof 2019 LA: used for training and evaluation. ESC-50: used for environmental noise in manipulation experiments.
Some description of the approach:
The paper first studies how simple audio manipulations can bypass existing audio deepfake detectors, then proposes CLAD, a contrastive-learning-based detector designed to be robust against those manipulations. CLAD trains an encoder to produce similar representations for differently manipulated versions of the same audio and dissimilar representations for different audio samples; it also adds length loss to cluster real audio features more closely.
Some description of the data:
The authors apply manipulation attacks to deepfake audio, including noise injection, volume control, fading, time stretching, resampling, time shifting, and echoes.
Keywords:
Contrastive learning; Audio deepfake detection; Manipulation attacks; Machine learning; Robust detection.
Instance Represent:
Each instance is an audio sample from ASVspoof 2019 LA. During CLAD pretraining, two manipulated/augmented versions of the same audio form a positive pair, while samples from different audio files form negative pairs. For model input, audio is repeated or clipped to a fixed length of 64,600 samples.
Dataset Characteristics:
The study uses the training set of ASVspoof 2019 LA for training and the evaluation set for testing. It excludes ASVspoof 2021 LA and DF because pretrained baseline models performed poorly on them.
Subject Area:
Audio deepfake detection, manipulation-attack robustness, contrastive learning, machine learning security, and robust forensic detection.
Associated Tools:
CLAD, AASIST, RawNet2, Res-TSSDNet, SAMO, MoCo-style contrastive learning, Adam optimizer, ESC-50, and the CLAD GitHub artifact repository.