Inferring access-control policy properties via machine learning

To ease the burden of implementing and maintaining access-control aspects in a system, a growing trend among developers is to write access-control policies in a specification language such as XACML and integrate the policies with applications through the use of a policy decision point (PDP). To assure that the specified polices reflect the expected ones, recent research has developed policy verification tools; however, their applications in practice are still limited, being constrained by the limited set of supported policy language features and the unavailability of policy properties. This paper presents a data-mining approach to the problem of verifying that expressed access-control policies reflect the true desires of the policy author. We developed a tool to investigate this approach by automatically generating requests, evaluating those requests to get responses, and applying machine learning on the request-response pairs to infer policy properties. These inferred properties facilitate the inspection of the policy behavior. We applied our tool on an access-control policy of a central grades repository system for a university. Our results show that machine learning algorithms can provide valuable insight into basic policy properties and help identify specific bug-exposing requests.