Overview
Healthcare IoT devices are increasingly targeted by cyber attacks. This project builds ML-based detection models that analyze network traffic patterns to identify anomalies and potential threats against IoMT infrastructure.
IoMT Protocol Analysis
Deep packet inspection is applied against DICOM traffic for medical imaging devices — parsed for abnormal association requests, unencrypted patient identifiers in PDU headers, and unexpected C-STORE/C-GET operations signalling data exfiltration. HL7 FHIR RESTful exchanges are monitored for irregular resource access patterns on Patient and Observation endpoints, bulk export triggers from unauthorized actors, and anomalous JWT usage. The MQTT layer — common in bedside monitors and infusion pumps — is analysed for topic enumeration attacks, malformed CONNECT packets, and command injection via retained messages. Proprietary medical device protocols are reverse-engineered through protocol fuzzing to build behavioural baselines.
Network Traffic Pipeline
PCAP ingestion feeds session reassembly via scapy and tshark, breaking bidirectional flows into 5-tuple conversations with configurable idle timeouts. Flow extraction yields per-session statistics: packet count, byte volumes, TCP flag distributions, and retransmission ratios. Feature engineering enriches flows with protocol-specific fields — DICOM Presentation Context IDs, FHIR endpoint paths, MQTT topic lengths and QoS levels — plus temporal features such as inter-arrival time (IAT) mean, variance, and percentiles. The result is a dense feature vector of ~180 dimensions.
Machine Learning Models
The detection stack employs a three-tier ensemble. Isolation Forest performs unsupervised anomaly detection by isolating outliers via recursive random partitioning — effective for zero-day attack discovery. LightGBM provides supervised classification using gradient-boosted decision trees optimised for the 1000:1 class imbalance inherent in medical traffic. A soft-voting ensemble combines their probability outputs with weights calibrated via Bayesian optimisation. Threshold tuning is guided by the F-beta score, balancing recall for high-severity events against false-positive rates in clinical environments.
Adversarial Evaluation
Model robustness is assessed via evasion attack simulations using FGSM and PGD against the LightGBM surrogate, constrained to perturbations realistic within medical network bounds (bounded latency shifts, plausible packet-size deltas). Resilience is quantified by the adversarial accuracy drop and minimum perturbation magnitude to flip a prediction. Current results show an 11% accuracy drop under PGD (ε=0.05), motivating ongoing adversarial training and feature squeezing defences.
Alerting Framework
When an anomaly exceeds the decision threshold, a severity score (1–10) is computed from anomaly magnitude, model confidence, asset criticality, and protocol risk tier. Alerts are serialised as Common Event Format (CEF) and forwarded via Syslog (RFC 5424) to Elastic Security, Splunk, and Wazuh SIEM platforms. Each alert carries flow metadata, triggered features, and the ensemble output vector for triage. A local alert buffer with exponential backoff deduplication prevents alert storms during burst attacks.
Impact
Contributes to securing critical healthcare infrastructure by enabling early threat detection before patient data or medical device integrity is compromised.