Graph Neural Networks for IoMT Intrusion Detection

Two related studies on intrusion detection for Internet of Medical Things (IoMT) networks using graph neural networks, evaluated on the CICIoMT2024 benchmark. The first, From Flow Features to Communication Topology, builds an adaptive graph attention network over device communication topology and introduces PCAP-level, session-aware validation; this exposed a 16 percentage-point accuracy gap versus the naive test-set early stopping used in prior work, and domain-typed edges improved macro-F1 by 7.4 points with a four-fold reduction in variance. The second, Temporal Graph Neural Networks for IoMT Intrusion Detection, combines a temporal GAT with a GRU and post-pooling temporal feature injection, reaching 0.853 macro-F1 versus 0.724 for a static baseline (+12.9 points) across five temporal granularities. Both manuscripts (2026) are co-authored with Prof. Abdulhamit Subasi at the University at Albany. Stack: Python, PyTorch Geometric, scikit-learn, SLURM on a DGX A100 cluster. Code: github.com/osmyildiz/iomt-graph-representation