Question 40
IIIWhich scenario best illustrates the privacy advantage of federated learning over traditional centralized training?
Correct answer: A
Explanation
Federated learning improves privacy because training happens on-device: "Model gradients are computed locally on user devices." Only "aggregated updates are sent to a central server," so the raw user data never leaves the device. That avoids the centralized model’s exposure of personal data during collection and training.
Why each option is right or wrong
A. Model gradients are computed locally on user devices, and only aggregated updates are sent to a central server, never the raw user data.
Under the privacy-by-design principle in GDPR Article 25, processing should minimize disclosure of personal data, and this scenario does so by keeping the underlying records on the device while transmitting only model updates. The key factual distinction is that the server receives only aggregated parameter changes, not the original user inputs, so there is no centralized pool of raw data exposed to collection, storage, or training on the server side.
B. All user data is periodically uploaded to a central server, trained on, and then deleted after model training.
Deleting later still requires raw data to be centralized first, which removes the core privacy advantage.
C. Only a small, random subset of user data is sent to the central server for training.
Sending less raw data is still centralized data collection, not federated learning’s local-training approach.
D. Data is fully centralized but pseudonymized before model training.
Pseudonymized centralized data still leaves user devices and can remain sensitive or re-identifiable.