The project aims to ensure the secure use of large language models (LLMs) in companies by developing the Differential Sensitivity Awareness (DSA) framework, which aims to adapt LLMs to manage security-critical data and differentiate access to information according to user permissions. Through a finely tuned identification and access control system, the framework provides legally compliant control over different “personas” within the same model to regulate access to sensitive information. These personas are also a focus of DFKI's research in the project, which aims to bring them to maturity primarily through the use of reinforcement learning with human feedback (RLHF), a technology that has recently gained prominence as part of natural language interaction systems in combination with generative AI.
To ensure broad use, all methods are developed across models and as open source, and are optimized specifically for the German-speaking market. The quality of the data set is crucial for training and evaluation, which is why its creation, i.e., a corpus of documents with different sensitivity characteristics, is another focus of the project. Furthermore, the technology is to be evaluated in a real-world environment, also with regard to resistance to attacks such as prompt injection. Finally, the subsequent use of the technology will be ensured by creating video tutorials and materials for a manual, which will also make it easier for users without AI expertise to get started.
Partners
L3S Research Center, DFKI, CISPA, Laverana