Scholarlink

<p>In recent years, facial forgery using Deepfake technology has become a significant threat to social security and privacy. The increasing visual quality of forged face images and the growing diversity of Deepfake models pose substantial challenges to the generalization and robustness of existing detection algorithms. In this paper, we present DSF, a <b>D</b>eepfake detection framework that leverages the complementary RGB <b>S</b>patial features and <b>F</b>requency-domain features. While single-domain features can capture forgery clues to some extent, they fail to fully leverage cross-domain complementarities, and simple feature fusion often produces suboptimal results. To address these limitations, DSF is designed to extract a comprehensive set of features, including RGB spatial, local high-frequency, and global high-frequency information. Moreover, we introduce a dedicated enhancement module to ensure effective integration of these diverse features. This module is composed of two primary mechanisms: (i) a self-enhancement block that highlights forgery clues within the spatial features, and (ii) mutual-enhancement blocks that facilitate cross-domain interaction between RGB and frequency-domain representations. Extensive experiments on public deepfake datasets validate the performance of our method. Moreover, the method maintains robustness against traditional image attacks, such as Gaussian noise.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deepfake detection method based on complementary enhancement of spatial-frequency domain features

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Kai Zhou,
  • Chenxu Wang,
  • Guanglu Sun,
  • Linsen Yu,
  • Jun Wang

摘要

In recent years, facial forgery using Deepfake technology has become a significant threat to social security and privacy. The increasing visual quality of forged face images and the growing diversity of Deepfake models pose substantial challenges to the generalization and robustness of existing detection algorithms. In this paper, we present DSF, a Deepfake detection framework that leverages the complementary RGB Spatial features and Frequency-domain features. While single-domain features can capture forgery clues to some extent, they fail to fully leverage cross-domain complementarities, and simple feature fusion often produces suboptimal results. To address these limitations, DSF is designed to extract a comprehensive set of features, including RGB spatial, local high-frequency, and global high-frequency information. Moreover, we introduce a dedicated enhancement module to ensure effective integration of these diverse features. This module is composed of two primary mechanisms: (i) a self-enhancement block that highlights forgery clues within the spatial features, and (ii) mutual-enhancement blocks that facilitate cross-domain interaction between RGB and frequency-domain representations. Extensive experiments on public deepfake datasets validate the performance of our method. Moreover, the method maintains robustness against traditional image attacks, such as Gaussian noise.