VAULT: Verified Access Control for LLM-Based Knowledge Graph Querying

Kurzfassung

The exponential growth of unstructured textual data in enterprise environments has made automated Knowledge graph generation essential for efficient information management. Although recent advances in natural language processing have enabled automated knowledge extraction, organizations face two critical challenges: maintaining domain specificity in knowledge representation and ensuring secure, role-based access to sensitive information. VAULT (Verified Access Control for Large Language Model (LLM)-Based Knowledge Graph Querying) presents a novel framework that combines ontology-driven knowledge extraction with dynamic access control mechanisms. The framework introduces three key innovations: (1) a configurable domaindriven node structure that enforces domain-specific Knowledge organization through semantic validation, (2) a multitiered access control mechanism that implements both document-level restrictions and node-level visibility patterns, and (3) an LLMpowered inference engine that dynamically filters Knowledge graph traversal based on user authorization levels. We implement our approach using a prototype system that demonstrates the automated conversion of natural language text into structured knowledge graphs while maintaining security constraints. Our experimental evaluation encompasses comprehensive testing across 16 different open-source LLMs, analyzing their Performance under varying access control conditions and authorization levels. The results demonstrate the framework’s effectiveness in maintaining information security while preserving query Response quality across different access tiers. The adaptability of the framework makes it particularly valuable for industries handling sensitive information, such as healthcare, finance, and intellectual property management, where both domain specificity and information security are paramount. This paper contributes to the field by bridging the gap between generic knowledge graph generation and domain-specific requirements while providing empirical evidence for the effectiveness of multilevel access control in LLM-based knowledge systems.