Paper finds learned access-control deny sets often unsound
Original title: Auditing the Hand-Written Inputs of Learned Access Control
Unchecked hard-deny clauses would block 16-78% of grants; certified sets withhold nothing, and hybrid ranking is one-class in all 24 cells.
An arXiv paper audits the hand-built shield and region map used with learned access control on five public ABAC Lab policies. Exhaustive checks find clauses that transcribe explicit restrictions unsound at rates of 0.50-0.70 across an author and three blind language-model elicitors, mostly because other rules still grant the request. Hybrid routing needs an abstaining two-class rule set that common ABAC miners do not emit. For additive losses, most selection rules give opposite verdicts on false-non-grant and false-permit, and a released artefact regenerates every reported number.
- Relations checked contain up to 794,250 tuples.
- Ninety-seven percent of counterexamples are granted by other rules.
- Shipped unchecked, deny sets would withhold 16-78% of policy grants.
- Ranking under a cap is one-class in all 24 cells for the two commonest keys.
- The only rule favourable on both losses holds for one policy and one learner setup.
From the original article
1 min read at arxiv.orgLearned access control is typically deployed behind two hand-built artefacts: a hard-deny set enforced by a small trusted shield, and a region map that decides which requests mined rules answer and which the learner answers. Verifying the shield or measuring the learner validates neither. We audit both on the five public ABAC Lab policies.
By Arsalan Jalil Sarghaleh · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.