Apprentissage fédéré
Définition
Federated learning trains models across many devices or organizations while keeping raw data local. Only model updates (par ex. gradients) are shared, reducing confidentialité and regulatory risk.
Utilisez-le quand data cannot be centralized (par ex. hospitals, phones) but you still want a shared machine learning model. Privacy is improved compared to sending raw data; additional techniques (differential confidentialité, secure aggregation) can be layered. See AI ethics for confidentialité and governance context.
Comment ça fonctionne
Le serveur détient le modèle global et l'envoie aux clients (appareils ou organisations). Chaque client s'entraîne localement sur sests own data and sends updates (gradients or model diff) back. The server aggregates updates (par ex. FedAvg: average the client models or gradients) and produces 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 confidentialité (updates can leak information; DP or secure aggregation mitigate).
Cas d'utilisation
Federated learning fits when data must stay on devices or silos and you still want a shared model.
- Training on sensitive data (par ex. healthcare, finance) without centralizing it
- Mobile and edge devices (par ex. keyboard suggestions, on-device ML)
- Cross-organization collaboration under confidentialité constraints