The proliferation of Internet of Things (IoT) devices has created an unprecedented attack surface, rendering traditional security measures inadequate. While AI-driven Intrusion Detection Systems (IDS) offer a promising defense, their conventional centralized architecture raises significant data privacy and scalability concerns. Federated Learning (FL) has emerged as a privacy-preserving alternative, enabling collaborative model training on decentralized data. However, the inherent structure of FL introduces a critical vulnerability: the global model is susceptible to poisoning attacks from malicious clients who manipulate their local updates. Most existing FL-based IDS frameworks fail to address this adversarial threat, limiting their real-world applicability. This paper introduces FedRobust-IDS, a novel FL framework engineered for adversarial-aware intrusion detection in IoT networks. FedRobust-IDS incorporates a robust aggregation mechanism that identifies and mitigates the impact of malicious model updates by performing anomaly detection on the submitted parameters. This approach preserves the privacy benefits of FL while significantly enhancing its security posture. We conduct a comprehensive evaluation using the large-scale, realistic CICIoT2023 dataset. The results demonstrate that under sophisticated model poisoning attacks, FedRobust-IDS maintains high detection accuracy and resilience, drastically outperforming standard Federated Averaging (FedAvg), which suffers a catastrophic performance collapse. This work represents a crucial step toward developing and deploying trustworthy, secure, and privacy-preserving AI for cybersecurity in adversarial environments