Which technique trains models locally on devices and aggregates updates?

• A. Differential privacy
• B. Secure multiparty computation
• C. Homomorphic encryption
• D. Federated learning

Answer: (D)

Federated learning trains models right on users’ devices using their local data, and then only the model updates are sent to a central server to be combined into a new global model. Each device learns from its own data, computes updates (like gradients or new parameters), sends those updates back, and the server aggregates them—often by a weighted average—before distributing the improved global model for another round. This setup directly matches the idea of local on-device training with update aggregation, while keeping raw data on devices and reducing what needs to be transferred.

Differential privacy adds noise to protect individual data in the outputs or updates, which can accompany federated learning but isn’t the mechanism that defines the on-device training and update aggregation. Secure multiparty computation and homomorphic encryption enable secure calculations, including secure aggregation, but the core concept described is specifically federated learning.