Leveraging High-Frequency Diversified Augmentation for general deepfake detection

Abstract

With the rapid advancement of deepfake technology, the visual quality of synthesized faces has significantly improved, raising serious security concerns about the misuse of facial manipulation techniques. As a result, deepfake detection has become a central focus within the multimedia forensics community. Recent studies have highlighted discrepancies between forged and genuine images in the high-frequency components. However, these studies have not fully addressed the inconsistency in high-frequency distributions across different datasets, which can lead to overfitting since models are trained on a limited range of high-frequency features. To overcome this challenge, we propose a High-Frequency Diversified Augmentation (HFDA) method designed to broaden the variation range of high-frequency features in training images. Specifically, our approach perturbs the amplitude spectra of the training data to generate augmented images with enhanced diversity in the high-frequency bands. Additionally, we introduce a forgery artifact consistency learning strategy to guide discriminative feature learning, aligning augmented images with their corresponding raw images. Extensive experiments demonstrate that the proposed HFDA method achieves superior or comparable performance to state-of-the-art methods across several widely used datasets. The code is available at https://github.com/laizhm/HFDA.