Jamming remains a significant threat to 5G network reliability and security, despite extensive research. This work addresses critical scalability and robustness gaps in previous approaches by introducing SHIELD: a scalable, holistic framework to evaluate jamming interference and support Machine Learning (ML)-based detection techniques without relying on costly external hardware. To validate our approach, we developed a realistic 5G testbed featuring a power-modulated jammer, commercial off-the-shelf Android devices, and a Software-Defined Radio (SDR)-based Radio Access Network (RAN). Our experiments demonstrate this setup generates complex interference patterns challenging prior detection methods. SHIELD’s novel methodology involves synchronously collecting native logs from User Equipment (UE) and Next-Generation Node B (gNB), capturing a comprehensive view of network behavior. It overcomes existing methods’ shortcomings against subtle, long-term interference by employing a robust preprocessing pipeline with interpolation and sliding-window aggregation for multi-layer feature extraction. We assess several lightweight classifiers (Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Gradient Boosting (GB), Random Forest (RForest)) across diverse real-world scenarios. Our evaluation reveals that while existing methods may exceed 90% accuracy in controlled settings, their performance can drop below 70% under varying conditions. In contrast, SHIELD’s log-based framework maintains approximately 94% accuracy on unseen data, offering a scalable, cost-effective, and robust solution for large-scale 5G deployments.