Proxy-Only Fog-Cloud Bidirectional Distillation for Privacy-Preserving, Deployable IoMT ECG Diagnosis under Cross-Site Heterogeneity

This work presents a proxy-only fog-cloud bidirectional knowledge distillation framework for privacy-preserving cross-site ECG diagnosis in Internet-of-Medical-Things (IoMT) systems. To satisfy data-governance constraints, hospital fog nodes and a cloud teacher collaborate solely via logits exchange on a public proxy dataset, avoiding patient data, labels, and model parameters. The proposed method stabilizes learning under cross-site heterogeneity and label shift while reducing communication overhead. Experiments on multi-site ECG data achieve 98.1% accuracy. An edge-oriented quantized FPGA deployment (1.98 ms latency and 0.169 mJ per inference) further demonstrates low-latency and energy-efficient real-time inference.