Enhancing Internet of Things Security: A Federated Learning‐Based Hybrid Model With Graph Neural Networks and Transformers

ABSTRACT The intrusion detection in Internet of Things data is the primary task to provide a robust security against the several ongoing attacks. In the conventional intrusion detection mechanisms, several security limitations are encountered due to greater latency and ineffective feature selection and a lack of graph‐structure utilization. So, to resolve these limitations, this work introduces a Federated Multi‐hop Attention Graph Transformer mechanism for improved intrusion detection in the Internet of Things applications. The transformer and multi‐hop attention graph neural networks are combined in the proposed model for superior intrusion detection. The proposed model ensures effective temporal learning from the sequence data by using a transformer. To understand the vector representation, the proposed model applies a multi‐hop attention graph neural network. For effective results, a fusion layer is applied to fuse representations of both components. The proposed system ensures improved privacy and scalability with federated learning. The utilization of federated learning avoids the need for raw data for cooperative model training across different devices. In the experimental results, four benchmark datasets are deployed for the proposed model's performance evaluation. For a fair performance comparison between the existing method and the proposed model, some evaluation metrics are deployed. The experimental findings demonstrate that the Federated Multi‐hop Attention Graph Transformer achieves improved accuracy and F1‐score of 98.21% and 97.27%, which is higher than the other related methods.