Family-FL: Privacy-Preserving ECG Monitoring on Sub-5KB Models
Researchers propose Family-Grouped Hierarchical Federated Learning (Family-FL) for ultra-resource-constrained wearables. The three-tier architecture uses the family as a privacy boundary for intra-family aggregation before global synchronization. A Tiny CNN-LSTM model with 669 parameters, INT8-quantized to 4.65KB Flash and 2.95KB RAM, fits STC32G12K128-class microcontrollers. Experiments on MIT-BIH Arrhythmia Database show feasibility for privacy-preserving ECG monitoring.
Key facts
- Family-FL uses family as a natural privacy boundary
- Model has only 669 parameters
- INT8 quantization reduces model to 4.65KB Flash and 2.95KB RAM
- Targets STC32G12K128-class microcontrollers
- Evaluated on MIT-BIH Arrhythmia Database
- Three-tier hierarchical architecture
- Privacy-preserving collaborative training
- Aims to detect arrhythmias via continuous ECG monitoring
Entities
—