In recent years, the incorporation of machine learning models within enterprise and cloud computing applications has introduced significant risks of model exposure from unauthorized feature access, abusive inferences, and other forms of exploitation. Current methods of access control rely mainly on role-based policies that are not dynamic enough to cater to emerging threats, changing user behavior, and context-dependent security policies. In this paper, we propose an AI-powered access control scheme for managing feature-level exposure in machine learning models and reducing the risk of model exposure to a minimum. Our approach leverages behavioral analysis and risk scoring capabilities to enforce access restrictions based on contextual analysis of feature-level usage and activities. In particular, our policy engine analyzes the interaction between users and model features, detects suspicious activity, and restricts access accordingly. Experiments carried out in a closed environment have shown that the proposed framework effectively limits unauthorized feature exposure while preserving the operational efficiency of deployed models.

Keywords
Reinforcement Learning, Autonomous incident response, Threat detection.