A novel intrusion detection system for IIoT in 5G networks using attention-augmented federated learning and lightweight transformer architectures.

The swift proliferation of Industrial Internet of Things (IIoT) devices within 5G networks has intensified cybersecurity vulnerabilities owing to their limited processing capabilities and fluctuating network conditions. This work presents an innovative intrusion detection system (IDS) designed for IIoT within 5G environments, overcoming the shortcomings of conventional hybrid deep learning models. Our methodology presents an attention-enhanced federated learning framework integrating a lightweight CNN (MobileNetV3) for feature extraction with a Mini-Transformer for sequence modeling, to attain superior accuracy and computational efficiency. A feature selection technique based on mutual information, integrated with self-attention layers, prioritizes essential network traffic features (e.g., packet size, flow time) while discarding redundant ones, hence improving detection accuracy. The federated learning framework enables decentralized training among IIoT devices with limited resources, ensuring both scalability and the maintenance of user privacy. Comprehensive evaluations conducted on the Edge-IIoTset and ToN-IoT datasets indicate that the proposed methodology yields a binary classification accuracy of 98.5% on the Edge-IIoTset, exceeding the performance of a standalone MobileNetV3 model by 2.1% and surpassing baseline models such as LSTM and VGG-19 by margins ranging from 3.5% to 6.3% under the same training parameters. Additionally, with regards to the ToN-IoT dataset, the model achieves an F1-score of 0.985, marking improvements of 2.0% over MobileNetV3-SVM and 6.5% over LSTM baseline models.