FEDERATED TINYML WITH POST-QUANTUM SECURITY: QUANTIFYING THE PRIVACY–ROBUSTNESS TRILEMMA IN CRITICAL CARE IOMT
Keywords:
Federated Learning, TinyML, Post-Quantum Cryptography (PQC), ML-KEM-512, IoMT Security, Non-IID Healthcare Data, Byzantine-Robust Aggregation, Explainable AI (XAI) for Critical CareAbstract
Edge intelligence has brought real-time, critical care monitoring into Internet of Medical Things (IoMT) for healthcare applications, but manifests a fundamental trilemma at the device level: how to balance patient privacy protection against gradient leakage with crypto resilience against quantum adversaries and robustness in learning from data under poisoning attacks—while achieving it all within stringent resource limits of TinyML microcontrollers. Current methods focus on these problems separately and do not provide an integrated framework that scores the relationships among them.
We propose Federated TinyML with Post-Quantum Security, an integrated edge-to-server architecture that jointly incorporates Federated Differential Privacy (FDP), Post-Quantum Cryptography (PQC), and a novel Robust Aggregation mechanism. The framework features three core contributions: (1) a lightweight Decision Tree ensemble optimized for ESP32-class microcontrollers, combining high inference efficiency with inherent explainability through aggregated feature importance; (2) a practical implementation of the NIST-standardized PQC scheme ML-KEM-512 (Kyber-512) to secure model updates during transmission, with minimal latency overhead; and (3) a Performance-Based Filtering strategy in the aggregation phase that autonomously identifies and discards poisoned client updates.
We evaluate the system on a clinician-validated, non-IID ICU dataset simulating decentralized data from three hospitals in Khyber Pakhtunkhwa, Pakistan. The global model achieves an overall accuracy of 95.64%, with an F1-score of 0.9935 for Status 3 (Critical)—a life-threatening condition requiring immediate intervention. Under a sustained Byzantine poisoning attack, the aggregator consistently rejects malicious updates, enabling the system to recover accuracy from 79.94% to 95.60% across training rounds. Explainability analysis confirms that the model prioritizes clinically meaningful features such as Respiratory Rate and Diastolic Blood Pressure.
To the best of our knowledge, this is first end-to-end IoMT framework to simultaneously address TinyML efficiency, post-quantum security, and robustness to adversarial data in critical care settings. It is establishing a foundational benchmark for secure, intelligent healthcare infrastructure.
