Fraunhofer IOSB model reliably detects deepfakes

Fraunhofer Institute of Optronics, System Technologies and Image Exploitation (Fraunhofer IOSB) has developed an improved approach to the detection of images created through deepfake technology.

The RealOrRender project, concluded in May 2026 and prepared by Fraunhofer IOSB in collaboration with the German Federal Office for Information Security (BSI), was launched to address the growing challenges posed by AI-generated image content, where advantages for creative applications or medical imaging sit alongside issues of disinformation and criminality.

RealOrRender takes a hybrid approach to deepfake detection, combining a conventional classification process based on deep learning with an evaluation of how well an image can be reconstructed using a generative model. If the reconstruction deviates significantly from the original image, this indicates that the image is real.

Combining these two approaches significantly increases detection accuracy, while ensuring that the system remains reliable even as generative models evolve and produce increasingly realistic results.

One of the project’s key concerns is transparency, commented Fraunhofer IOSB, and rather than simply classifying an image as AI-generated, the system should explain why it arrived at a decision. This involves various attribution-based and segment-based explainable AI (XAI) processes to visualize the image features and regions on which the decision is based. 

Hybrid XAI approach outperforms current methods

"First, we use an AI image generator to reconstruct the image," said Fraunhofer's Andreas Specker. "Then an AI model takes over the classification and carries out a hybrid calculation of the reconstruction error. We finally obtain an estimate displaying recognition accuracy as a percentage. The overall detection rate ranges between 85 and 91 percent, although it can certainly be higher in individual cases."

One key aspect of RealOrRended is its use of XAI, an approach in which the reasoning behind an AI model's decisions is made more understandable and transparent than might be the case if no such conditions were placed on the model. Fraunhofer HHI previously made use of explainable AI in its work on the early detection of skin cancer, as one example of the approach.

"We use XAI methods to highlight the image areas and structures that contribute to the model's decision," noted Nadia Burkart, Head of the Applied Explainable AI group at Fraunhofer IOSB. "It can thus be clearly understood why an image is classified as real or AI-generated. For example, distinctive textures or characteristic frequency patterns can hint at the synthetic origin of an image."

According to Fraunhofer IOSB data, researchers use XAI methods that can be classified into heatmap-based and segment-based approaches based on their explanation output formats. Heatmap-based approaches visualize which areas of the image contributed most to the model's decision and the degree to which these factors influenced the evaluation. Segment-based methods analyze contiguous image segments and show which semantic regions contributed to the model's decision.

In tests on a large image dataset, the researchers were able to demonstrate that this hybrid, explanation-oriented approach outperforms current methods. The research findings are now being incorporated in a demonstrator that will initially assist the BSI in detecting deepfakes.

"The new hybrid approach significantly improves detection accuracy, while explainable AI processes ensure transparent results," commented the project.