Fraunhofer Develops Tool for Reliable Deepfake Image Detection
Researchers have created a new hybrid method that reliably detects AI-generated images and explains the reasoning behind the classification. The tool significantly enhances accuracy and transparency.
Researchers at Fraunhofer have developed a new tool, RealorRender, capable of detecting AI-generated deepfake images and explaining the classification process. This hybrid approach combines traditional deep learning with reconstruction analysis from generative models to significantly improve detection accuracy.
The method works by reconstructing an image using AI and then analyzing the reconstruction error. The better an image can be reconstructed, the more likely it is to be AI-generated. Detection success rates range between 85% and 91%, with explainable AI (XAI) methods highlighting the specific image areas and structures that contributed to the classification.
The project aims to address the growing risk of misinformation posed by realistic AI-generated images. By integrating detection with explainability, researchers seek to create more transparent and trustworthy systems to combat the spread of manipulated content.
The developed method was tested on a large image dataset and demonstrated superior performance compared to existing techniques. The findings are being incorporated into a demonstrator tool to assist the German Federal Office for Information Security (BSI) in identifying deepfakes, providing users with detection results and detailed explanations.