Abstract
With rapid advances in computing systems, there is an increasing demand for more effective and efficient access control (AC) approaches. Recently, Attribute Based Access Control (ABAC) approaches have been shown to be promising in fulfilling the AC needs of such emerging complex computing environments. An ABAC model grants access to a requester based on attributes of entities in a system and an authorization policy; however, its generality and flexibility come with a higher cost. Further, increasing complexities of organizational systems and the need for federated accesses to their resources make the task of AC enforcement and management much more challenging. In this paper, we propose an adaptive ABAC policy learning approach to automate the authorization management task. We model ABAC policy learning as a reinforcement learning problem. In particular, we propose a contextual bandit system, in which an authorization engine adapts an ABAC model through a feedback control loop; it relies on interacting with users/administrators of the system to receive their feedback that assists the model in making authorization decisions. We propose four methods for initializing the learning model and a planning approach based on attribute value hierarchy to accelerate the learning process. We focus on developing an adaptive ABAC policy learning model for a home IoT environment as a running example. We evaluate our proposed approach over real and synthetic data. We consider both complete and sparse datasets in our evaluations. Our experimental results show that the proposed approach achieves performance that is comparable to ones based on supervised learning in many scenarios and even outperforms them in several situations.
Introduction
Managing access control in complex environments, such as Home IoT, often leads to a "rule explosion" where the number of possible attribute combinations grows exponentially. While Attribute-Based Access Control (ABAC) is the gold standard for flexibility—granting access based on user, object, and environmental attributes—its manual management is error-prone and lacks adaptivity. Traditional Supervised Learning (SL) can mine policies from logs but requires massive amounts of pre-labeled data and fails to adapt when system dynamics change. This paper introduces a framework that automates ABAC policy management using Reinforcement Learning (RL) to create a system that is both adaptive and data-efficient.
The ABAC-RL Framework
We treat ABAC policy learning as a Contextual Bandit problem, a specific form of Reinforcement Learning. The system operates through a continuous feedback loop between an Authorization Engine and its environment.
In this model:
- The Agent: The Authorization Engine (AE) which maintains the policy $\pi$.
- The State ($s_t$): A vector of attributes including the Subject (user), Object, Environment, and the requested Operation.
- The Action ($a_t$): A binary decision—Permit or Deny.
- The Reward ($r_t$): Feedback from the system administrator or resource owner.
Reward Mechanism
The agent learns by comparing its decision ($d_{AE}$) against the owner's intended decision ($d_w$). If the decisions match (True Positive or True Negative), the agent receives a positive reward; if they conflict (False Positive or False Negative), it receives a penalty.
Key Innovations: Initialization and Planning
To overcome the "cold start" problem inherent in RL, we propose four initialization methods: utilizing general AC rules, default user settings, default capability actions (e.g., "deny" for sensitive locks), and mining historical logs.
Additionally, we introduce a planning approach based on Attribute Value Hierarchies (AVH). By understanding the relationships between attributes (e.g., a "teenager" is a type of "minor"), the agent can propagate learned permissions across the hierarchy. This allows the model to make educated guesses about states it hasn't explicitly visited, significantly accelerating the learning process in sparse datasets.
Experimental Results
We evaluated the framework using the Amazon Kaggle Dataset (33,000 real-world access requests) and synthetic Home IoT Datasets.
- RL vs. Supervised Learning: The RL approach achieved a loss rate of ~5.8% on the Amazon dataset, nearly identical to the 5.5% achieved by a fully supervised model, despite RL learning on the fly without upfront labels.
- The Power of Planning: Implementing AVH-based planning reduced the Progressive Validation Loss (PVL) by 10% to 25% across various scenarios.
- Adaptability to Policy Shift: The agent demonstrated the ability to handle changes in rules mid-stream. Specifically, the Online Cover algorithm proved most effective at balancing exploration and exploitation, recovering from policy shifts faster than traditional methods.
Conclusion
ABAC-RL offers a promising path toward automated, adaptive access control. By leveraging contextual bandits and domain-specific hierarchies, the framework achieves performance comparable to supervised learning while maintaining the flexibility to adapt to evolving organizational needs and user preferences.
Summary updated