Friday 14 March 2025
A new approach to detecting deepfakes has been developed, and it’s a game-changer for identifying manipulated images and videos.
Deepfakes are AI-generated fake media that can be used to spread misinformation or even deceive people into thinking they’re seeing something real. They’re often created using generative adversarial networks (GANs) and diffusion models, which can produce incredibly realistic results.
However, detecting these fakes has been a major challenge. Traditional methods rely on visual features such as texture, color, and lighting, but these can be easily manipulated by the creators of deepfakes. As a result, many fake images and videos have slipped through the net, causing concern about their potential impact on society.
The new approach, developed by researchers at an engineering research center in China, takes a different tack. Instead of focusing on visual features, it extracts information from both local and global frequency domains to detect forgeries.
The method, known as LoSFB (Local Spatial-Frequency Based), uses two core modules: Local Spatial Frequency Feature Extractor (LoSFE) and Global Frequency Domain Feature Extractor (GloFDE). The LoSFE module captures local spatial-frequency information by using a combination of discrete wavelet transform (DWT) and sliding window tiling. This allows it to extract detailed and subtle texture features that are often missed by traditional methods.
Meanwhile, the GloFDE module extracts global frequency domain information from the phase component of the fast Fourier transform (FFT). This provides a broader view of the image or video, allowing the model to detect patterns and anomalies that may not be apparent at the local level.
The combination of these two modules enables LoSFB to capture a wider range of features and improve its ability to detect deepfakes. In tests on 34 diverse generative models, LoSFB outperformed existing methods by a significant margin, achieving an accuracy rate of over 84%.
One of the key advantages of LoSFB is its ability to generalize across different types of deepfakes. While many detection methods are limited to specific types of forgeries, LoSFB can detect a wide range of fakes, including those generated using GANs and diffusion models.
The implications of this technology are significant. With LoSFB, it may be possible to develop more robust systems for detecting deepfakes, which could help mitigate the spread of misinformation and protect people from online deception.
Cite this article: “Breakthrough Technology Unlocks New Way to Detect Deepfakes”, The Science Archive, 2025.
Deepfakes, Ai-Generated, Fake Media, Gans, Diffusion Models, Image Detection, Video Detection, Losfb, Feature Extraction, Frequency Domain
