Access control policy mining, which extracts or infers access control policies from existing system logs or configurations within a given environment, is widely applied in policy migration. No policy mining algorithm can produce completely accurate policies, necessitating human intervention to correct errors before implementation. However, humans may fail to revise the produced policies without fully understanding all permission details. In this paper, we propose an on-the-fly framework to assist human revisions by mining usable policies. This framework is formulated as an approximate optimization problem, designed to avoid permission errors and redundancy by balancing submodularity and modular costs through an iterative search for access rules. The search space is guided by two pruning techniques: a tight optimistic estimate to only eliminate unpromising candidates and a queue cutter to sample promising candidates in advance. Experimental evaluations on two real-world and three publicly available synthetic datasets indicate that: 1) our method produces more concise results than existing methods, achieving a 93.2% reduction in redundancy; 2) our method is at least five times faster than state-of-the-art approaches. To further validate the usability of the policies obtained by our approach, we conducted a user study involving 30 participants and 7 large language models (LLMs). The results show that 90.9% of participants, including LLMs such as gpt-4o-mini, successfully modified our mined policies to meet given permission goals.
Read full abstract