FL-TSA-TabNet: Federated Tabular Network for Intrusion Detection

The rise of cyber-physical systems and IoT centers have heightened the need to have intrusion detection systems (IDS) based on scalability, privacy-preservation, and explainability. Although models like CNNs and BiLSTMs work effectively as deep learning models, they can be difficult to interpret and can experience limitations associated with centralized data. In order to overcome these challenges, this work proposes FL-TSA-TabNet, a new federated intrusion detection model that incorporates Temporal Self-Attention into the interpretable TabNet architecture. The model uses a Dual-Stage Hybrid Selector (DSHS) with a correlation-aware reliefF and SHAP-based ranking as the core feature of optimal feature relevance and non-redundancy. With federated learning, the model also allows decentralized training among the distributed nodes and data privacy. On the CSE-CIC-IDS2018 dataset, FL-TSA-TabNet performed with an accuracy of 97.83%, better than the traditional (Random Forest, XGBoost) and the hybrid deep models (CNN-GRU, ResNet-BiLSTM). It also exhibited excellent adversarial robustness, quick inference speed, and low model complexity, which makes it incredibly appropriate in edge deployment. This work sets a new benchmark in intrusion detection by fusing explainable learning, temporal modeling, and privacy-aware federated training, paving the way for next-generation IDS in smart networks and critical infrastructure environments.