A robust hybrid deep learning framework for IoT intrusion detection with adaptive feature embedding and...

The rapid expansion of Internet of Things (IoT) ecosystems has increased exposure to cyber threats, highlighting the need for robust and generalizable intrusion detection systems (IDS). Existing approaches often suffer from weak feature representation, poor handling of class imbalance, and inconsistent evaluation practices, limiting their real-world applicability. To address these issues, this study proposes a hybrid IDS framework that combines transformer-based feature encoding, a Temporal Fusion Transformer (TFT) for advanced sequential pattern modeling, and a lightweight Convolutional Neural Network (CNN) for spatial feature extraction, integrated through a confidence-driven adaptive fusion mechanism. The methodology includes Adaptive Synthetic Sampling (ADASYN)-based imbalance handling, mutual information-based feature selection, and contrastive self-supervised embedding to enhance representation quality. The model is evaluated on benchmark datasets, including CIC-IDS2017 and BoT-IoT, using stratified k-fold cross-validation and metrics such as accuracy, precision, recall, F1-score, and ROC-AUC. Quantitatively, the proposed hybrid IDS achieved an accuracy of 98.3%, precision of 97.8%, recall of 96.9%, F1-score of 97.3%, and ROC-AUC of 98.0%, while maintaining a cross-dataset accuracy of 95.4% when evaluated across the CIC-IDS2017 and BoT-IoT datasets. The proposed framework offers a reliable and interpretable solution for IoT intrusion detection, with future work focusing on real-time deployment and adaptive security mechanisms.