Technical Summary: RuleConEx

Problem Statement

Deep learning-based access control (DLBAC) models, such as those proposed by Nobi et al., offer high accuracy in predicting access decisions but suffer from complete opacity. This lack of interpretability creates a significant legal and operational barrier, particularly regarding compliance with regulations like the GDPR (Article 22), which mandates the ability to explain automated decisions.

Existing solutions for explainability in this domain are insufficient:

• Post-hoc methods (e.g., SHAP, LIME) provide local approximations that are not faithful to the actual decision mechanism and can be manipulated.
• Separate self-explaining models address only one aspect of explainability. HyperLogic generates global IF-THEN rules but offers no actionable counterfactuals for denied users. HyConEx generates local counterfactuals but provides no global auditable rules for administrators.
• Current limitations: No existing architecture intrinsically unifies global symbolic rules and local actionable counterfactuals within a single forward pass, nor do they ensure that counterfactuals are valid categorical tuples by construction.

Methodology

The authors propose RuleConEx (Rule-based Counterfactual eXplainer), an intrinsically interpretable neural network that unifies logical rules and counterfactual explanations.

Architecture

RuleConEx is built around a shared hypernetwork that parameterizes two distinct branches, fusing their outputs into a single decision:

1. Shared Hypernetwork: A TabResNet trunk that generates parameters for the branches. It is queried by:
   • A global query (learned vector + Gaussian noise) to parameterize the rule branch, ensuring a single, consistent rule set for all requests.
   • A local query (latent encoding of the input instance) to parameterize the counterfactual branch.
2. Rule Branch (Global): A binarized DR-Net (Differentiable Rule Network) with $K=48$ rule neurons. It produces a set of IF-THEN rules decoded directly from the network weights. The rules use a specific vocabulary of relational literals (e.g., $user\_attr_i = resource\_attr_j$) and negations, aligning with the ABAC policy language of Xu and Stoller.
3. Counterfactual Branch (Local): A local linear classifier parameterized by the hypernetwork. It generates counterfactuals by computing the gain of modifying categorical attributes to reach a target class. Crucially, it enforces validity by construction, ensuring the output is a valid categorical tuple and respecting a mask of immutable attributes (e.g., resource attributes).
4. Fusion: The logits from both branches are combined via a weighted sum with learned, softmax-normalized weights ($\alpha_r$ and $\alpha_h$) to produce the final access decision.

Training and Loss Function

The model is trained jointly to minimize a composite loss function:
$$\mathcal{L} = \mathcal{L}_{CE} + \lambda_r \mathcal{L}_{rule} + \lambda_h \mathcal{L}_{loc} + \lambda_{cf} \mathcal{L}_{cf} + \lambda_{sp} \mathcal{L}_{sparsity} + \lambda_{len} \mathcal{L}_{length} + \lambda_{ent} \mathcal{L}_{entropy}$$
Key components include:

• Counterfactual Loss ($\lambda_{cf}$): Encourages the model to classify modified tuples (counterfactuals) into the target class.
• Rule Constraints: Penalizes rule length (max 8 literals) and encourages single-class voting (entropy).
• Faithfulness Measurement: Unlike previous works that assume faithfulness, RuleConEx measures the strict and weighted agreement between the decoded rules and the final fused decision.

Key Contributions

1. Unified Architecture: RuleConEx is the first model to produce an access decision, global IF-THEN rules, and local counterfactuals in a single forward pass.
2. Intrinsic Validity and Faithfulness:
   • Rules are exact: The function computed by the rule branch is fully decodable, and its faithfulness is measured rather than assumed.
   • Counterfactuals are valid categorical tuples by construction, handling discrete attributes without post-hoc projection.
3. Relational Literals: The model explicitly incorporates relational constraints (equality between user and resource attributes) into the input space and rule vocabulary. The authors demonstrate that without these, symbolic models (like decision trees) fail to capture the policy structure effectively.
4. Tunable Trade-off: The paper identifies a single coefficient, $\lambda_{cf}$ (the weight of the counterfactual loss), as a control mechanism. Sweeping this coefficient allows for a continuous shift between:
   • User-oriented explainability: Shorter counterfactuals (fewer attribute changes).
   • Auditor-oriented explainability: Higher consistency between rules and counterfactuals (more counterfactuals justified by a global rule).
   • This trade-off is achieved without statistically detectable differences in accuracy.

Experimental Results

The model was evaluated on 11 DLBAC benchmark datasets (3 real-world Amazon datasets and 8 synthetic datasets) using 5 random seeds.

• Accuracy: RuleConEx achieved a mean accuracy of 92.1%. This is statistically indistinguishable from HyConEx, Random Forests, and Linear SVMs, and only 0.6 percentage points lower than an opaque MLP baseline.
• Rule Quality:
  • Rules cover 80% of requests on average.
  • Strict faithfulness (agreement between rules and final decision) is 0.80.
  • Rules are concise (mean length ~4.5 literals) and primarily relational.
• Counterfactual Quality:
  • 96% of counterfactuals are valid categorical tuples.
  • They modify an average of 2 attributes.
  • Flip success rate is 100%.
• Consistency: The consistency between rules and counterfactuals (fraction of valid counterfactuals justified by a target-class rule) ranges from 0.18 (default setting) to 0.43 (when $\lambda_{cf}$ is relaxed). This is significantly higher than the consistency achieved by juxtaposing separately trained HyperLogic and HyConEx models (0.22).
• Ablation Studies:
  • The shared hypernetwork itself is neutral; the benefits come from joint training.
  • Gaussian noise in the global query is superfluous.
  • Relational literals are critical; removing them causes a significant drop in rule accuracy and coverage.

Significance and Claims

The paper claims that its primary contribution is not a gain in predictive accuracy, but the unification of two complementary forms of explainability within a single, tunable model.

• Explicit Control: RuleConEx provides an explicit tuning mechanism ($\lambda_{cf}$) to balance the needs of the denied user (who needs a short, actionable path to access) and the auditor (who needs global, consistent rules). Two juxtaposed models cannot offer this continuous trade-off.
• Security Implications: The authors argue that exposing both rules and counterfactuals creates a new attack surface. They propose a privilege separation model:
  • Auditors receive global rules and branch logits.
  • Requesters receive only the decision and a single, restricted counterfactual (limited to mutable attributes) upon denial.
  • This limits the information leakage regarding the policy boundary compared to systems that return full counterfactual sets for all classes.

The authors conclude that while the model achieves high accuracy and interpretability, the consistency between rules and counterfactuals is a tunable property rather than a structural guarantee, and future work is needed to quantify policy leakage and validate the utility of these explanations with human auditors.