Large Language Models for Automated Data Access Policy Creation in Data Spaces

Ryan Ford, Rita Beigaite, Arash Hajikhani

Abstract: Data spaces are being developed to enable more standardized and secure data sharing. They provide a framework where data can be exchanged following predefined access control policies. However, manually defining machine-readable policies can be a time-consuming and error-prone task. It poses challenges for the adoption and scalability of Data Spaces. In this paper, we investigate the automatic creation of machine-readable data access policies that could support users of Data Spaces. We assess the ability of Large Language Models to generate structured data access policies and validate their adherence to the defined ontology. Our findings reveal that while Large Language Models excel at producing syntactically valid policies (98% accuracy) and maintaining ontological compliance (90% accuracy), they fundamentally struggle with encoding complex logical relationships in access control rules, with only 1% of generated policies passing logical consistency validation, highlighting the continued necessity of human expertise in policy creation.

External IDs:doi:10.1007/978-3-032-04878-3_4