RuleConEx: an intrinsically interpretable neural network unifying logical rules and counterfactual explanations for data access control

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Deep learning based access control models such as DLBAC are accurate but opaque, which makes them difficult to audit against the accountability requirements of regulations such as the GDPR. Existing self-explaining models provide only one kind of explanation: global rules (HyperLogic) or local counterfactuals (HyConEx). We propose RuleConEx, a model trained jointly around a shared hypernetwork that produces, in a single forward pass, an access decision, IF-THEN rules decoded exactly from the network weights and expressed in the language of ABAC policies, and counterfactuals that are valid categorical tuples by construction. Rule faithfulness and counterfactual validity are measured rather than assumed. On the eleven DLBAC benchmarks, evaluated with five seeds, a baseline juxtaposing separately trained HyperLogic and HyConEx models, and statistical tests with Nemenyi and Holm corrections, RuleConEx reaches 92.1% mean accuracy, statistically indistinguishable from HyConEx, random forests and linear SVM, and 0.6 points below an opaque MLP. Its rules cover 80% of the requests with a strict faithfulness of 0.80, and 96% of its counterfactuals are valid, changing two attributes on average. Sweeping the weight of the counterfactual loss shows that this single coefficient continuously shifts the trade-off between user-oriented explainability (counterfactuals of 1.9 attributes) and auditor-oriented explainability, where the consistency between rules and counterfactuals rises from 0.18 to 0.43, with no statistically detectable difference in accuracy. RuleConEx therefore provides an explicit tuning mechanism for this trade-off, which two juxtaposed models do not offer.