Adaptive graph attention-based federated learning for IoT intrusion detection: mitigating poisoning attacks

The widespread adoption of Internet of Things (IoT) networks introduces critical security challenges, particularly due to poisoning attacks targeting federated learning (FL)-based intrusion detection systems (IDSs). Traditional FL methods, such as FedAvg, are vulnerable to adversarial updates, which compromise model integrity and reliability. To address these limitations, this article proposes an Adaptive Graph Attention-Based Federated Learning (AGAT-FL) framework designed to enhance the resilience of IoT-based IDSs. AGAT-FL combines dynamic trust-aware aggregation using graph attention networks (GAT), a hybrid convolutional neural network-gated recurrent unit (CNN-GRU) deep learning model for spatial-temporal anomaly detection, and Mahalanobis distance-based filtering to identify and suppress adversarial contributions. Trust scores are adaptively assigned to participating clients based on historical performance and behavioral indicators, allowing AGAT-FL to downweight suspicious updates while preserving data privacy. Experimental evaluations on two benchmark IoT security datasets, N-BaIoT and CIC-ToN-IoT, demonstrate that AGAT-FL consistently outperforms state-of-the-art FL methods. It achieves up to 94.01% accuracy, 93.50% precision, 94.00% recall, and 93.75% F1-score on N-BaIoT, and 91.02% accuracy, 91.00% precision, 91.30% recall, and 91.15% F1-score on CIC-ToN-IoT. Additionally, the use of explainable AI techniques such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) enhances transparency by identifying key features contributing to anomaly classification. These results underscore AGAT-FL as a robust, interpretable, and scalable solution for securing FL-based IoT networks against sophisticated poisoning attacks.