The rapid evolution of Internet of Things (IoT) networks has presented serious security threats because of the enormous volume of distributed data produced by connected devices. Traditional IDSs (IDS) usually follow centralized data collection, resulting in communication overhead, scalability issues, and privacy problems. Although federated learning (FL) offers a way to train distributed models while preserving privacy, many current FL-based AD techniques cannot be adapted to account for the interaction relationships between IoT devices. To address these challenges, this study introduces the federated graph attention network (FL-GAT) for anomaly detection in IoT-edge environments. The proposed framework treats IoT devices as graph nodes and introduces a multi-head graph attention mechanism to capture the spatial interaction among devices while guaranteeing data privacy by adopting federated learning. Local models are trained in a distributed manner on edge devices without sharing raw data. Distributed IoT attack scenarios were used to evaluate the proposed framework using the TON_IoT and Bot-IoT benchmark datasets. Experimental results show that FL-GAT achieved 95.2 % accuracy and 94.5 % F1 score on TON_IoT and 94.8 % accuracy and 94.1 % F1 score on Bot-IoT, with better results than centralized deep learning and federated deep learning baseline models, and graph-based baseline models. Furthermore, the attention mechanism enhances the interpretability of the model by identifying the key interactions between devices that lead to unusual activities. Although the proposed framework shows encouraging performance and scalability, the evaluation was conducted using benchmark IoT datasets under a simulated experimental setting. Future work will focus on real-world deployment scenarios, dynamic network conditions, and lightweight edge optimization for resource-constrained IoT devices.