Our PhD Student Andreas Müller under the supervision of Erwin Quiring and Asja Fischer successfully published his paper titled „Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models“ at the top-tier Conference CVPR taking place in Nashville (USA) from June 11 to June 15, 2025!

This work introduces multiple new attacks against the current state-of-the-art in imperceptible watermarking for detecting AI-generated images, so-called semantic watermarks. This calls their trustworthiness into question and prompts further research into robust watermarking. Key Highlights:
– Watermark Forgery & Removal: The forgery attack transfers a watermark from a generated image to any other real image. The removal attack deletes a watermark in a generated image.
– Imperceptible: The visual changes made to an image are not noticeable to humans because they target subtle visual features.
– Black-Box: Due to a surprising transferability property of AI-image generators, the attacker does not need access to the original closed-source AI-model and can instead use any other open-source model.
– Single-Example: While comparable approaches need thousands of watermarked reference images, this attack outperforms them with just a single example at hand.

Check out the Preprint here!