Which privacy risks does federated learning reduce, and which remain?

Instruction: Explain how keeping raw data local changes exposure. Discuss leakage from updates and models, secure aggregation, differential privacy and the trust assumptions required for any privacy claim.

Context: Assess the privacy boundaries of federated learning without assuming that local data storage guarantees confidentiality or regulatory compliance.

Updated

Example Answer

Federated learning improves privacy mainly by keeping raw training data on the client instead of centralizing it. That reduces direct exposure and often aligns better with regulatory or institutional constraints.

But I would be careful with the wording "ensure." Federated learning alone does not guarantee privacy, because updates can still leak information. Real privacy protection usually requires additional controls such as secure aggregation, differential privacy, access restrictions, and careful system logging.

Common Poor Answer

A weak answer says federated learning is private because data never leaves the device, ignoring leakage through gradients or updates.

Related Questions