Adversarial Deepfake Example Detection based on Phase Spectrum Image

• Shaomei Li,
• Ruiyang Huang,
• Fan Zhang,
• Hongxin Zhi,

Vol. 19, No. 6, June 30, 2025
10.3837/tiis.2025.06.003, Download Paper (Free):

• Adversarial deepfake example
• Adversarial example
• Black-box Attack
• Deepfake detection
• Phase spectrum image
• White-box attack

Abstract

With the advancement of deepfake technology, deepfake videos, particularly those involving face swapping, have rapidly spread across the internet, posing significant threats to national security and social stability. Consequently, deepfake detection techniques have emerged. Existing research has revealed that deep neural network (DNN)-based detection models are vulnerable to adversarial examples, enabling malicious spreaders to circumvent detection systems using adversarial example techniques. Our research discovers that although deepfakes generated through adversarial attacks are challenging to detect in the spatial domain, they leave traces in the image’s phase spectrum. Building upon this finding, we propose an approach for detecting adversarial deepfake samples based on phase spectrum images. Firstly, we extract the phase spectrum from facial images, then reconstruct the face’s phase spectrum image. Finally, we train an image classifier using the phase spectrum images to detect adversarial deepfake samples. Our method is straightforward yet effective and can be integrated into a joint detection framework to achieve high detection rates.

Statistics

Cite this article