Cyber-Physical Systems (CPS) integrate Internet of Things (IoT) technology to enable seamless communication and control across interconnected devices. However, IoT-driven CPS faces critical security challenges, including data breaches and denial-of-service attacks. Machine learning and deep learning techniques in Centralized Learning (CL) and Federated Learning (FL) settings are used for Intrusion Detection Systems (IDS), but the non-IID nature of IoT data complicates both approaches. Non-IID data, including data skew, affects model performance and convergence, with FL showing promise but requiring further exploration. This study investigates the performance of CL and FL in IDS for IoT-driven CPS under IID (balanced) and non-IID (imbalanced) scenarios. We compare both approaches using metrics such as accuracy, precision, recall, and F1-score. Our findings reveal that FL outperforms CL in IID settings, achieving higher accuracy. Moreover, FL’s resilience is demonstrated under non-IID conditions, outperforming CL by up to 1.3% in certain cases. Given the decentralized and heterogeneous nature of IoT-driven CPS, FL proves to be a robust and scalable solution for intrusion detection.