beyond rules: a learned model for nuanced access control in semantic search

Rule-based access control systems in semantic search are often too rigid for the nuanced nature of information retrieval, frequently blocking access to highly relevant documents that fall just outside a user s strict permissions. this paper moves beyond rules by proposing a novel approach where a neural network learns a nuanced access control policy from data. i trained a multi-layer perceptron (mlp) on a synthetically generated dataset that encodes a flexible security policy based on semantic similarity, user level, and document level. the results demonstrate that the learned model can make intelligent, context-aware decisions, successfully identifying and assigning a high permission score to relevant documents that a rigid system would block, while still enforcing security boundaries. this work shows the feasibility of replacing hand-crafted rules with a data-driven model, offering a path towards more adaptive and intelligent security for modern information retrieval systems.