Project funding: Improving the detection of deepfakes
Researchers develop prototype that identifies fake images generated by AI
Deepfakes are spreading rapidly and are increasingly difficult to detect. An IT forensics group at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) is currently working on a tool in conjunction with secunet Security Networks AG that detects images generated by AI automatically and reliably. The project has received 350,000 euros of funding from the Federal Agency for Breakthrough Innovation SPRIN-D.
Angela Merkel is chatting with Vladimir Putin in a beer garden. The Pope is a DJ at a mixing desk. Tom Cruise is locked in an embrace with Paris Hilton. These images are circulating on the Internet and receive millions of clicks, but they are very realistic fakes known as deepfakes. “Deepfake generators are becoming more and more powerful and are also readily accessible,” says PD Dr. Christian Riess, Head of the Multimedia Security research group at the Chair of Computer Science 1 (IT Security Infrastructures). “This means images and videos with manipulated contents are being spread very quickly and they seem increasingly genuine.” At best, they are just entertaining. However, the fakes can cause political and social controversy if they are not detected as such.
A fundamental new approach
A project at FAU is currently tackling this very problem. In conjunction with secunet Security Networks AG, the IT specialists in Erlangen are developing a universal prototype that should reliably detect deepfakes created by various AI generators. “Our approach is fundamentally different from other methods,” explains Sandra Bergmann, a doctoral candidate in the Riess working group. “Image detection programs are usually trained by presenting them with a large number of examples. The software then learns to classify the images and to differentiate between genuine images and those produced by AI.”
![Graphic shows how an image is created with the camera and how an AI-generated image is created and what the traces that can be detected in an AI image look like.](https://www.fau.eu/files/2025/01/Bild_Pressemitteilung_ges-300x159.jpg)
The prototype developed by FAU and secunet also uses image classification. In addition, it also uses large pre-trained neural networks from AI generators in order to extract image characteristics. “The advantage is that we can use our tool to check images that have been created by previously unknown deepfake generators,” says Bergmann. “This means the tool can react spontaneously without the need for previously training the detector with thousands of pieces of data.” The aim of the FAU researchers is to merge as many detectors and data traces as possible, thus making the prototype robust against errors.
At the same time, secunet is finding ways of implementing the tool in existing digital infrastructures and making it detect deepfakes on social media platforms reliably. “The high reliability needed for detecting deepfakes is only one requirement for the solution. The detection system must also be able to make decisions quickly, even if it receives several inquiries simultaneously from an application,” explains Dr. Benjamin Tams, project manager at secunet.
What is SPRIN-D?
SPRIN-D is an initiative of the Federal Ministry of Education and Research and considers itself an incubator for breakthrough innovations in Germany and Europe. It identifies, validates, finances and manages projects that show potential as breakthrough innovations. In contrast to financing projects that focus solely on research, SPRIN-D aims to provide support for new marketable products, technologies, business models and services that change the lives of as many people as possible in the long term. The researchers from FAU are using their many years of experience in the field of deepfake detection and their specialist expertise in AI-supported image compression to develop the prototype.
Further information:
PD Dr. Christian Riess
Chair of Computer Science 1 (IT Infrastructures)
Phone: + 49 9131 85 69919
christian.riess@fau.de