en.Wedoany.com Reported - Germany's Federal Office for Information Security (BSI, Bundesamt für Sicherheit in der Informationstechnik) and the Fraunhofer Institute of Optronics, System Technologies and Image Exploitation (Fraunhofer IOSB) have jointly developed a deepfake detection method called RealOrRender, aimed at identifying AI-generated fake images and making the detection process traceable. Currently, the realism of AI-generated images and videos continues to improve, with large language models now capable of producing anatomically correct limbs, realistic shadow projections, and even simulated heartbeats, making it increasingly difficult to distinguish between real footage and deepfakes. Planned tightening of EU laws may not fundamentally change this situation, making technical solutions a research priority.
The BSI notes that automated detection methods typically only operate reliably under specific conditions. To address this, the BSI and Fraunhofer IOSB collaborated to develop the RealOrRender method. This approach employs a two-part process: first, an AI image generator reconstructs the input image, then an AI model classifies it and calculates a mixed reconstruction error. Andreas Specker, Senior Scientist at the Fraunhofer IOSB Video Assistance for Security and Assistance Systems research group, explains that this process can identify whether an image is a deepfake.
The specific mechanism of RealOrRender is as follows: after receiving an image, the system uses a pre-trained diffusion model to perform reverse processing, generating a mathematical fingerprint (noise map) of the image and a reconstruction result, i.e., intentionally re-forging the original image. The system then scans the image for errors. The logic of this method is: if the original image is a deepfake, its mathematical fingerprint will be similar to RealOrRender's reconstruction result; if it is a real photo, more reconstruction errors will occur because real photos contain more imprecision and natural noise. Using a training dataset containing approximately 120,000 images from 18 image generators, the system calculates an estimated deviation and provides an assessment of whether the image is a deepfake. According to researchers, the detection performance ranges between 85% and 91%.
RealOrRender also integrates an explainable AI (XAI) component, which generates a heatmap at the end of the analysis, marking details in the original image that point to forgery, including features such as faces, hair, hands, and objects in the background. This helps users understand which deepfake indicators the system has identified and enhances trust in the model and its results. It is reported that RealOrRender has already been deployed as a demonstrator within the BSI.
