Core Concepts
By blending frequency knowledge from fake faces into real faces, the proposed FreqBlender method can generate pseudo-fake faces that closely resemble the distribution of wild fake faces, enhancing the learning of generic forgery features for DeepFake detection.
Abstract
The paper introduces a new method called FreqBlender to generate synthetic fake faces, known as pseudo-fake faces, by blending frequency knowledge. Existing methods typically generate these faces by blending real or fake faces in the color space, but they overlook the simulation of frequency distribution, limiting the learning of generic forgery traces.
To address this, the authors propose a Frequency Parsing Network (FPNet) that can adaptively partition the frequency space into three components: semantic information, structural information, and noise information. They hypothesize that the forgery traces are likely hidden in the structural information. By blending the structural information of fake faces with real faces, FreqBlender can generate pseudo-fake faces that closely resemble the distribution of wild fake faces in the frequency space.
Since there is no ground truth for the frequency distribution, the authors design dedicated training objectives that leverage the inner correlations among different frequency components to instruct the learning process of FPNet.
Extensive experiments on multiple DeepFake datasets demonstrate the effectiveness of FreqBlender in enhancing DeepFake detection performance, outperforming state-of-the-art methods. The method can also complement existing spatial-blending techniques, making it a potential plug-and-play strategy for other detection approaches.
Stats
The frequency range of forgery traces varies across different fake faces due to its high dependence on face content.
Forgery traces may not be concentrated on a single frequency range but could be an aggregation of various portions across multiple ranges.
Quotes
"By blending the structural information of fake faces with real faces, FreqBlender can generate pseudo-fake faces that closely resemble the distribution of wild fake faces in the frequency space."
"Since there is no ground truth for the frequency distribution, the authors design dedicated training objectives that leverage the inner correlations among different frequency components to instruct the learning process of FPNet."