Föderiertes Lernen
Definition
Föderiertes Lernen trainiert Modelle über viele Geräte oder Organisationen wobei die Rohdaten lokal bleiben. Only model updates (z. B. gradients) are shared, reducing Datenschutz and regulatory risk.
Verwenden Sie es, wenn data cannot be centralized (z. B. hospitals, phones) but you still want a shared machine learning model. Privacy is improved compared to sending raw data; additional techniques (differential Datenschutz, secure aggregation) can be layered. See AI ethics for Datenschutz and governance context.
Funktionsweise
Der Server hält das globale Modell und sendet es an Clients (Geräte oder Organisationen). Jeder Client trainiert lokal auf seinents own data and sends updates (gradients or model diff) back. The server aggregates updates (z. B. FedAvg: average the client models or gradients) and erzeugt a new global model, then broadcasts again. Rounds repeat until convergence. Challenges: heterogeneity (non-IID data, different compute), communication cost (limit round count or update size), and Datenschutz (updates can leak information; DP or secure aggregation mitigate).
Anwendungsfälle
Federated learning passt, wenn data must stay on devices or silos and you still want a shared model.
- Training on sensitive data (z. B. healthcare, finance) without centralizing it
- Mobile and edge devices (z. B. keyboard suggestions, on-device ML)
- Cross-organization collaboration under Datenschutz constraints