A Database-Independent LLM Framework for Real-Time Authorization in Retrieval-Augmented Generation

Keywords: Retrieval-Augmented Generation, Access Control, LLM Authorization, Multi-modal Security, Real-time Authorization, Trustworthy AI Systems

TL;DR: A database-independent framework for RAG access control achieving 93% accuracy in real-time authorization across text and image-based security without manual preprocessing in multi-organizational environments.

Abstract: Retrieval-Augmented Generation (RAG) systems have become effective tools for improving language model capabilities by integrating external knowledge. However, these systems face significant security problems when handling sensitive information, particularly concerning unauthorized data access. This paper presents a new access-controlled RAG framework that includes dynamic security measures without needing manual database construction or extensive preprocessing of user-specific document collections. Our approach uses three specific knowledge bases and advanced language models to make real-time access control decisions. Experimental testing on a dataset of 167 posters from six organizations—University of Antwerp, Flanders Make, University of Ghent, University of Hasselt, KU Leuven, and VUB—shows 93\% overall accuracy in access control decisions, with 92\% accuracy for text-based control and 78\% for image-based control. The framework enables quick deployment of secure RAG systems while maintaining efficiency and responsiveness.

Submission Number: 68