RealOrRender

Background

With their ability to create deceptively realistic images, generative AI models offer a wide range of possibilities, e.g. for creative applications or medical imaging. However, the associated risks are increasingly coming under scrutiny, particularly when these artificial images are deliberately used for disinformation. The aim of the RealOrRender project is to develop processes with which AI-generated images can be reliably and robustly distinguished from real photos.

Hybrid approach to deepfake detection

RealOrRender takes a hybrid approach: A conventional classification process based on deep learning is combined with a method that evaluates 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.

Transparency through explainable AI

One of the project’s key concerns is transparency: Rather than simply classifying an image as “AI-generated,” the system should explain why it arrived at this decision. This is achieved by deploying various attribution-based and segment-based explainable AI processes to visualize the image features and regions on which the decision is based. Evaluations performed using a large, purpose-built and diverse imaging dataset have shown that this hybrid, explanation-oriented approach clearly outperforms many processes currently used in research. RealOrRender accordingly provides effective technological protection against image manipulation and establishes a robust foundation for the responsible use of generative AI content.