This paper presents the design and generation of a novel high-fdelity intrusion detection dataset specifcally targeting 5G core control-plane attacks. The dataset is constructed using an Open5GSbased testbed integrated with my5G-RANTester, enabling realistic simulation of benign UE registration and advanced authentication-layer attacks, including MAC failure, SQN desynchronization, replay, brute force, NAS message manipulation, and denial-of-service scenarios. From raw packet captures, 25 protocol-aware features are engineered, combining flow-level statistics with entropy-based and sequence-consistency indicators that reflect 5G-AKA signaling logic. To validate the dataset’s effectiveness, multiple machine learning models—ranging from Decision Trees to ensemble methods such as Random Forest and XGBoost—are evaluated using Accuracy, F1-score, and cross-validation metrics under class imbalance conditions. Experimental results demonstrate that ensemble models achieve near-perfect classifcation performance with strong generalization capability, highlighting the discriminative power of semantic-aware features. The fndings confrm that context-aware feature engineering is essential for reliable intrusion detection in virtualized 5G core infrastructures