Aprendizaje federado
Definición
El aprendizaje federado entrena modelos a través de muchos dispositivos u organizaciones while keeping raw data local. Only model updates (por ej. gradients) are shared, reducing privacidad and regulatory risk.
Úselo cuando data cannot be centralized (por ej. hospitals, phones) but you still want a shared machine learning model. Privacy is improved compared to sending raw data; additional techniques (differential privacidad, secure aggregation) can be layered. See AI ethics for privacidad and governance context.
Cómo funciona
El servidor mantiene el modelo global y lo envía a los clientes (dispositivos u organizaciones). Cada cliente entrena localmente en susts own data and sends updates (gradients or model diff) back. The server aggregates updates (por ej. FedAvg: average the client models or gradients) and produce 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 privacidad (updates can leak information; DP or secure aggregation mitigate).
Casos de uso
Federated learning fits when data must stay on devices or silos and you still want a shared model.
- Training on sensitive data (por ej. healthcare, finance) without centralizing it
- Mobile and edge devices (por ej. keyboard suggestions, on-device ML)
- Cross-organization collaboration under privacidad constraints