TL;DR
This paper introduces FS-SEA, an efficient and extensible algorithm for automatically generating Relationship-Based Access Control (ReBAC) policies. By incorporating a novel neural network-based feature selection phase, FS-SEA significantly reduces the search space for evolutionary algorithms, leading to faster policy mining and higher quality policies compared to previous methods.
What Problem Does This Paper Address?
Migrating from legacy access control systems to more flexible and expressive frameworks like Relationship-Based Access Control (ReBAC) can be a significant undertaking. ReBAC policies allow for complex access rules based on relationships between entities (e.g., a user's relationship to a document's owner), offering greater expressiveness than traditional attribute-based systems. However, automatically generating these policies from existing data (policy mining) is challenging due to the exponentially large search space of possible policies. Existing ReBAC policy mining algorithms, while promising, are often limited in the complexity of policies they can handle and can be computationally expensive. This paper addresses the need for a ReBAC policy mining algorithm that is not only efficient and effective but also easily extensible to support new policy language features.
Methods
The core contribution of this paper is the FS-SEA* algorithm, which combines Feature Selection (FS) with a Simplified Evolutionary Algorithm (SEA).
- Feature Selection (FS): This novel phase uses a neural network (NN) to identify "useful" atomic conditions and constraints that are likely to appear in the target ReBAC policy. The NN is trained on SRA-tuples (subject, resource, action) labeled as permitted or denied. By analyzing the NN's learned weights, the algorithm quantifies the importance of each feature. Features are ranked based on their contribution to the NN's decision-making, and the top-ranked features are selected. This phase is crucial for reducing the search space for the subsequent evolutionary algorithm.
- Simplified Evolutionary Algorithm (SEA): This phase is a modified version of a previously proposed evolutionary algorithm (EA). The simplification involves removing the "rule generalization" step, making it easier to extend. SEA then uses the "useful features" identified by the FS phase to construct candidate ReBAC rules.
- Iterative Approach (FS-SEA*): The algorithm operates in an outer loop. In each iteration, it performs feature selection and then runs SEA to mine a single best rule for each (subject type, resource type, action) tuple, focusing on uncovered SRA-tuples. This iterative process allows the algorithm to progressively refine the policy and discover rules that might be missed in a single pass.
The paper also introduces two new constraint operators (subseteq and set equality) to demonstrate the extensibility of their approach.
Datasets
The paper uses a combination of datasets for evaluation:
- Realistic Case Studies: The authors use two case study policies from previous work: the "e-document" case study and the "workforce management" case study. These are based on real-world organizations but use synthetic object models.
- Synthetic Policies: To evaluate the algorithm's performance with the newly introduced operators and to explore larger, more complex scenarios, the authors developed a synthetic policy generator. This generator creates class models, object models, and authorization sets (AU) that mimic realistic policy structures while allowing for controlled complexity and the inclusion of new operators.
Key Insights and Ideas
- Feature Selection as a Bottleneck Reducer: The most significant insight is that a well-chosen feature selection phase can dramatically improve the efficiency and effectiveness of evolutionary policy mining. By pre-filtering features, the algorithm avoids wasting computational resources exploring irrelevant parts of the vast search space.
- Neural Networks for Feature Importance: Using neural networks for feature selection is a powerful approach. Their ability to implicitly learn complex interactions between features makes them well-suited for identifying relevant conditions and constraints in access control policies.
- Extensibility through Simplification: Simplifying the evolutionary algorithm by removing the rule generalization step is key to making the overall approach extensible. This allows new policy language features to be incorporated with minimal changes to the core mining logic.
- Iterative Refinement for Completeness: The iterative nature of FS-SEA* is crucial for ensuring that all necessary rules are discovered, especially for complex policies where some rules might only cover a small subset of authorizations.
Key Results
- Efficiency: FS-SEA was significantly faster than the baseline EA on larger synthetic policies, with running times reduced by orders of magnitude in some cases. For example, on one large object model, FS-SEA took about 2.7 hours, while SEA was aborted after 23 hours.
- Effectiveness: FS-SEA* produced policies with higher syntactic and per-rule semantic similarity to the original policies compared to SEA. It also generated policies with significantly lower Weighted Structural Complexity (WSC), indicating more concise and manageable policies.
- Comparability to Baseline: On smaller, realistic case studies, FS-SEA* achieved comparable effectiveness to the original EA in terms of policy similarity, demonstrating that the simplification did not harm performance on simpler problems.
- Extensibility Demonstrated: The algorithm successfully incorporated and mined policies with new constraint operators (
subseteqand set equality), showcasing its extensibility.
Limitations and Future Work
- Limitations:
- The paper acknowledges that on smaller search spaces (like those in the realistic case studies), the overhead of repeated feature selection in FS-SEA* can make it slightly slower than EA.
- The current implementation does not support inheritance between classes, which is a feature of some real-world policies.
- Mining policies with DENY rules is noted as future work.
- Future Work:
- The authors plan to incorporate inheritance into their algorithm.
- Extending the algorithm to mine policies that include DENY rules is a key area for future research.
- Further exploration of different feature selection techniques and NN architectures could yield additional improvements.
Summary updated