Search for: All records

Access control policies are crucial in securing data in information systems. Unfortunately, often times, such policies are poorly documented, and gaps between their specification and implementation prevent the system users, and even its developers, from understanding the overall enforced policy of a system. To tackle this problem, we propose the first of its kind systematic approach for learning the enforced authorizations from a target system by interacting with and observing it as a black box. The black-box view of the target system provides the advantage of learning its overall access control policy without dealing with its internal design complexities. Furthermore, compared to the previous literature on policy mining and policy inference, we avoid exhaustive exploration of the authorization space by minimizing our observations. We focus on learning relationship-based access control (ReBAC) policy, and show how we can construct a deterministic finite automaton (DFA) to formally characterize such an enforced policy. We theoretically analyze our proposed learning approach by studying its termination, correctness, and complexity. Furthermore, we conduct extensive experimental analysis based on realistic application scenarios to establish its cost, quality of learning, and scalability in practice. 

Mining algorithms for relationship-based access control policies produce policies composed of relationship-based patterns that justify the input authorizations according to a given system graph. The correct functioning of a policy mining algorithm is typically tested based on experimental evaluations, in each of which the miner is presented with a set of authorizations and a system graph, and is expected to produce the corresponding ground truth policy. In this paper, we propose formal properties that must exist between the system graph and the ground truth policy in an evaluation test so that the miner is challenged to produce the exact ground truth policy. We show that failure to verify these properties in the experiment leads to inadequate evaluation, i.e., not truly testing whether the miner can handle the complexity of the ground truth policy. We also argue that following these properties would provide a computational advantage in the evaluations. We propose algorithms to identify and correct violations of these properties in system graphs. We also present our observations regarding these properties and their enforcement using a set of experimental studies.