The exponential expansion of the Internet of Things (IoT) ecosystem has accelerated the need for real-time, distributed data analytics while intensifying privacy and security risks arising from centralized data collection. Federated Learning (FL) provides a promising alternative by collaboratively training global models across edge devices without exposing raw data. Nevertheless, conventional FL frameworks face multiple challenges, including susceptibility to gradient inversion, membership inference, and poisoning attacks, as well as significant communication and energy overheads on resource-constrained IoT nodes. To address these limitations, FL-ISM is proposed, a federated learning–based IoT security model that integrates secure aggregation, calibrated differential privacy, and Byzantine-resilient optimization with reputation-aware client selection and communication compression mechanisms. The system and threat model are formally defined, privacy and robustness guarantees are derived, and FL-ISM is evaluated on intrusion and anomaly detection benchmarks under non-IID data conditions. Experimental results demonstrate that FL-ISM not only achieves competitive predictive performance but also reduces uplink traffic and effectively mitigates backdoor and inference attacks, thereby enabling scalable, privacy-preserving, and secure analytics in safety-critical IoT environments.