Federated Intrusion Detection Using TabTransformer-TCN-BiGRU-Attention: A High-Accuracy Hybrid Deep Learning Approach

Abstract

In the context of Industry 4.0, safeguarding Industrial IoT (IIoT) networks against increasingly sophisticated cyber threats remains a critical challenge, as traditional Intrusion Detection Systems (IDSs) often struggle with scalability, adaptability, and data privacy concerns. This study addresses these limitations by introducing a novel hybrid deep learning architecture for anomaly-based intrusion detection in IIoT environments. The proposed model combines TabTransformer for contextual feature extraction, Temporal Convolutional Networks (TCN) and Bi-directional GRU (BiGRU) for temporal sequence modeling, and an attention mechanism to enhance focus on subtle attack patterns. Using the IoTID20 dataset, the model was first evaluated in centralized training, where it outperformed baseline models (LSTM, CNN, CNN-BiGRU, BiGRU) with an F1-score of 99.8%, accuracy of 99.6%, and an AUC of 0.999. To ensure privacy-preserving and communication-efficient deployment, the model was further implemented in a Federated Learning (FL) setting using Flower, enabling collaborative training across distributed clients without sharing raw data, and significantly reducing bandwidth consumption by exchanging only model parameters. Overall, the proposed approach contributes a scalable, accurate, and privacy-aware intrusion detection framework, positioning hybrid transformer-temporal architectures as promising solutions for secure and intelligent IIoT systems.