Agency for leapfrog innovations seeks radical solutions against deepfakes

The Federal Agency for Jump Innovations (Sprind) wants to strengthen the fight against highly realistic synthetic content, usually generated using artificial intelligence (AI). To this end, it recently launched an innovation competition on behalf of the Federal Ministry for Digital and Transport (BMDV) to identify and prevent such deepfakes. An expert jury selected a total of twelve teams from almost 50 applications, as Sprind announced on Tuesday. The participants will each receive up to 350,000 euros for their further work over the next seven months. Stage 2 will begin in May, in which a maximum of ten remaining teams will receive up to 375,000 euros each for the further development of their solution, which should be ready in November 2025.

In the 21st century, the integrity of digital media and information is increasingly threatened by deepfakes, explains Sprind. This technology poses "significant risks of disinformation, fraud and manipulation of public opinion". Thanks to scientific advances, there are already new findings and methods in the field of machine learning to develop more powerful deepfake detection systems. However, these have so far often been limited to niche applications or specific types of deepfakes. Moreover, available approaches are usually not reliable. Another shortcoming: "Prevention measures have so far been limited and cannot be seamlessly integrated into existing digital infrastructures."

Entire digital media chain to be covered

According to the tender, the prototype sought in the competition should now "demonstrate how image deepfakes can be reliably detected and authenticated". The entire system must be able to continuously adapt to new technologies. At the end of the process, at least three different use cases such as social media, news portals or video conferencing systems should be covered. "Scalability and adaptability to different digital platforms" are also required. Sprind competition manager Jano Costard explains: the aim is to consider the "entire value chain of digital media processing". This includes preventative authentication and securing of original content as well as real-time analysis and detection of deepfakes.

The teams selected by scientists, industry experts and investors come from research institutions, universities and start-ups in Germany, France and Slovenia. One of them is a group of researchers from the University of Nuremberg and the IT security company Secunet, which is using "multi-detectors" with "VeriTrue". This should make it possible to reliably classify genuine and fake images from any deepfake generator, even if they were not seen during training. The ReaLGuard team from the CeMOS research center in Mannheim is working on a universal detection method that is also based on content-independent features and adapts to new image generation methods. DeepShield from Secublox and the Bundeswehr University, for example, swears by forgery-proof watermarks and "blockchain-based verification".

(olb)