Description
Federated learning trains a global model across many clients (rooms, devices, sites) by exchanging gradient or weight updates without sharing raw data. It is attractive for CSI sensing because raw CSI is privacy-sensitive (it can leak room layouts and behaviour), and because most realistic deployments have many partial sites rather than one big labelled corpus.
When it's used
- Multi-site CSI sensing without centralising raw data
- Edge-device collaborative learning under privacy constraints
- Continual cross-deployment refinement of a shared CSI backbone
Limitations
- Non-IID client data degrades convergence
- Communication-cost overheads
- Privacy leakage from gradients still possible