HDFedAtt-IIoT: A Novel Privacy-Preserving Hybrid Deep Federated Learning Framework with Attention and Proximal Regularization for IIoT Systems

The growth of Industrial Internet of Things (IIoT) networks has made the need for intrusion detection systems (IDS) that are both secure and scalable even greater. This study presents HDFedAtt-IIoT, a hybrid deep federated attention-based framework that combines sophisticated feature selection, imbalance management, and adaptive federated optimization to enhance cyber-attack detection in IIoT networks. The first step in the process is to use Boruta to choose the most useful 30 features from the large X-IIoTID dataset. The SMOTE-Tomek method was used to fix uneven class distributions, making sure that all 19 attack categories were represented equally. The processed data was then sent to five federated clients, where they used an improved version of the Adaptive FedProx algorithm with dynamic regularisation to train their own models. The main framework architecture used convolutional layers, bi-directional GRUs, and an attention mechanism to find hybrid temporal-spatial dependencies in network traffic. The experimental evaluation demonstrated that the results quickly converged, with macro-F1 reaching 0.99 after 50 communication rounds, yielding a final accuracy of 99.08%, precision of 99.09%, recall of 99.07%, and macro-F1 of 99.06%. Class-wise analysis showed that the system was strong against a wide range of attacks, with rare classes being especially sensitive. Embedding visualisations through PCA and t-SNE showed that the framework could learn feature spaces that were very different from each other. Attention-based interpretability helped us understand which features were most important for making predictions. These results show that HDFedAtt-IIoT is a scalable, communication-efficient, and easy-to-understand IDS solution for protecting the next generation of IIoT ecosystems.