An explainable multi-head attention network for healthcare IoT threat detection based on the MedDefender-MHAN...

Area of Science:

• Cybersecurity
• Artificial Intelligence in Healthcare
• Internet of Medical Things (IoMT)

Background:

• Healthcare environments face critical cybersecurity risks due to the proliferation of Internet of Medical Things (IoMT) devices.
• Existing deep learning intrusion detection systems (IDS) lack the explainability required for clinical adoption and regulatory compliance (e.g., GDPR, FDA).

Purpose of the Study:

• To present MedDefender-MHAN, an explainable multi-head attention network for healthcare IoT threat detection.
• To embed interpretability directly into the network architecture for real-time explanations, unlike post-hoc methods.

Main Methods:

• A novel dual-stream architecture combining Convolutional Neural Networks (CNNs) and Transformer-based encoders.
• Integration of interpretability within the multi-head attention mechanism for gradient-weighted explanation generation.
• Evaluation on CICIDS2017 and TON_IoT benchmark datasets.

Main Results:

• Achieved high detection accuracies of 99.47% (CICIDS2017) and 98.92% (TON_IoT).
• Demonstrated sub-3ms inference latency and a throughput of 435 samples per second.
• Explainability evaluation showed 94.6% alignment with expert signatures and 91.9% temporal accuracy, outperforming post-hoc methods.

Conclusions:

• MedDefender-MHAN provides a clinically viable and regulatory-compliant security solution for healthcare IoMT infrastructure.
• The framework addresses the need for trustworthy AI-driven security with methodological transparency and clinical impact.
• It offers a practical approach to enhancing cybersecurity in regulated healthcare settings.