.png){kind=icon}
.png){kind=icon}
Return
Enhancing robustness in deepfake detection: a contrastive invariance approach
.png){kind=icon}
PRE
AI
H
N
M
P
DOI：10.1007/s00530-026-02590-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deepfake technology, which leverages artificial intelligence to create counterfeit audio and visual content, poses significant threats to identity verification and societal security. Traditional detection methods, primarily based on deep neural networks, are vulnerable to adversarial attacks. This paper introduces a detection framework based on noise-invariant contrastive learning to enhance adversarial robustness. By aligning feature representations of original and noise-augmented samples, our approach encourages the model to focus on semantic content while filtering out noise. Experiments on the FF++ dataset demonstrate that our method effectively resists various adversarial attacks, achieving over 90% accuracy in detecting adversarial Deepfake samples. This approach significantly outperforms conventional adversarial training, showing stronger generalization and a better balance between robustness and clean data performance.
Keywords:
Deepfake
Adversarial attack
Contrastive learning
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers