FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge

Hanzhe Li,Yuezun Li,Jiaran Zhou,Bin Li,Junyu Dong

2024-05-06

Abstract:Generating synthetic fake faces, known as pseudo-fake faces, is an effective way to improve the generalization of DeepFake detection. Existing methods typically generate these faces by blending real or fake faces in color space. While these methods have shown promise, they overlook the simulation of frequency distribution in pseudo-fake faces, limiting the learning of generic forgery traces in-depth. To address this, this paper introduces {\em FreqBlender}, a new method that can generate pseudo-fake faces by blending frequency knowledge. Specifically, we investigate the major frequency components and propose a Frequency Parsing Network to adaptively partition frequency components related to forgery traces. Then we blend this frequency knowledge from fake faces into real faces to generate pseudo-fake faces. Since there is no ground truth for frequency components, we describe a dedicated training strategy by leveraging the inner correlations among different frequency knowledge to instruct the learning process. Experimental results demonstrate the effectiveness of our method in enhancing DeepFake detection, making it a potential plug-and-play strategy for other methods.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

### Problems Addressed by the Paper

The paper aims to enhance the performance of DeepFake detection by generating pseudo-fake faces through the integration of frequency knowledge. Specifically, the paper addresses the following issues:

1. **Limitations of Existing Methods**: Existing pseudo-fake face generation methods mainly focus on the fusion in the color space, neglecting the distribution of pseudo-fake faces in the frequency space. This leads to models failing to learn generalizable forgery features.

2. **Utilization of Frequency Space**: The paper proposes a new method called FreqBlender, which analyzes frequency components and extracts structural information related to forgery traces. This frequency knowledge is then fused from fake faces into real faces to generate pseudo-fake faces that are closer to the real distribution.

3. **Adaptive Partitioning Network**: To overcome the issue of the frequency component distribution being non-fixed and potentially spanning multiple bands, the paper introduces an Adaptive Partitioning Network (Frequency Parsing Network, FPNet). This network dynamically partitions the frequency space and extracts semantic information, structural information, and noise information from it.

Through the aforementioned methods, the paper aims to generate pseudo-fake faces that are closer to the real forgery distribution, thereby enhancing the generalization ability of DeepFake detection models. Experimental results show that this method can significantly improve detection performance across different datasets and has the potential to serve as a plugin for existing methods.