Goals and approach
The “KI-FOGGER” project, funded by the BMFTR, aims to secure AI models so that they only function correctly with a special release. To achieve this, the team is specifically using methods that were originally developed for attacks on AI – for example, inserting false data (data poisoning) or installing secret access points (backdoors). Without appropriate authorization, the models only deliver distorted, unusable results. In this way, the knowledge contained in the model remains protected. The methods are then tested in real application scenarios in industry.
Innovations and prospects
The approach is novel: security is built directly into the model so that sensitive information remains protected even if the model itself is lost. This strengthens the technological independence of German companies and enables the secure and responsible use of AI. The project thus makes an important contribution to digital sovereignty.
