Edge-based access-control systems must evaluate identity, context, and policy constraints quickly enough to protect distributed resources without imposing unnecessary authentication. This study compares CatBoost and a one-dimensional convolutional neural network (1D CNN) as complementary models for contextual risk prediction in an enhanced access-control framework. Structured access records were encoded using login time, device trust, known location, biometric confidence, resource sensitivity, and Chinese Wall conflict status. CatBoost supplied interpretable risk estimates, whereas the 1D CNN classified access states as failed, uncertain, or verified. The 1D CNN achieved 84.08% test accuracy on 157 records; the verified class produced 96.77% recall and a 90.91% F1-score. System-level results further showed that adding biometric evidence and policy checks moved some context-only grants to step-up or deny decisions. The findings indicate that neither model is sufficient as a standalone security mechanism. A layered edge pipeline that combines structured risk scoring, contextual classification, identity confidence, and conflict-of-interest enforcement provides a more defensible basis for real-time grant, step-up, and deny decisions. Read more