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

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• The proliferation of Internet of Things (IoT) devices necessitates advanced cybersecurity measures.
• Existing intrusion detection systems (IDS) face challenges with feature representation, class imbalance, and evaluation consistency.
• These limitations hinder the practical application of IDS in complex IoT environments.

Purpose of the Study:

• To develop a robust and generalizable hybrid IDS framework for enhanced IoT security.
• To address limitations in existing IDS, including weak feature representation and class imbalance.
• To improve the reliability and interpretability of intrusion detection in IoT ecosystems.

Main Methods:

• A hybrid IDS framework integrating Transformer-based encoding, Temporal Fusion Transformer (TFT), and Convolutional Neural Network (CNN).
• Employed Adaptive Synthetic Sampling (ADASYN) for imbalance handling and mutual information for feature selection.
• Utilized contrastive self-supervised embedding for superior representation quality and a confidence-driven adaptive fusion mechanism.

Main Results:

• Achieved high performance on benchmark datasets (CIC-IDS2017, BoT-IoT) with 98.3% accuracy, 97.8% precision, 96.9% recall, 97.3% F1-score, and 98.0% ROC-AUC.
• Demonstrated strong cross-dataset generalizability with 95.4% accuracy.
• The proposed hybrid IDS framework provides a reliable and interpretable solution for IoT intrusion detection.

Conclusions:

• The developed hybrid IDS framework effectively addresses key challenges in IoT intrusion detection.
• The integration of advanced deep learning models and data handling techniques yields superior performance.
• Future research will focus on real-time deployment and adaptive security capabilities for evolving cyber threats.