The proliferation of the Internet of Things (IoT) has introduced unprecedented security challenges, demanding robust and scalable anomaly detection systems. Traditional centralized machine learning models are untenable due to severe privacy, scalability, and communication constraints. While Federated Learning (FL) has emerged as a promising privacy-preserving alternative, it struggles with the statistical heterogeneity of data (non-IID), high communication overhead, and static aggregation logic. This paper introduces Federated Swarm Intelligence (FSI), a novel hierarchical framework that applies Swarm Intelligence (SI) at two distinct levels to create a resilient and efficient collective intelligence system. At the edge, FSI employs a decentralized, ant-based clustering algorithm to autonomously group devices with similar data distributions, directly mitigating the negative impacts of non-IID data. At the core, it uses a Particle Swarm Optimization (PSO) algorithm at the server to intelligently aggregate models from cluster representatives, optimizing knowledge fusion. This dual-level SI optimization drastically reduces communication overhead by limiting server interactions to elected cluster heads and produces a more accurate global model. We conduct a rigorous empirical evaluation on the comprehensive CIC-IoT-2023 dataset, demonstrating that FSI significantly outperforms standard FL baselines in accuracy, communication cost, and convergence speed. This work establishes FSI as a viable architectural blueprint for building adaptive and self-organizing security solutions for the next generation of decentralized AI systems.