What is federated learning, and how does it apply to deep learning?
Instruction: Explain the concept of federated learning and its implications for training deep models.
Updated
Example Answer
Federated learning trains a shared model while participating devices or organizations keep their raw training records locally. In a common setup, the server sends model weights, selected clients train locally, and the server aggregates their updates.
Keeping raw records local reduces one form of data sharing, but it does not guarantee privacy. Gradients or model updates can reveal training information. Secure aggregation can hide individual updates from an aggregator under its protocol assumptions; differential privacy can limit information leakage under a specified privacy definition and budget. Neither is a substitute for access controls and a clear threat model.
I would also plan for non-identical client data, uneven participation, unreliable connections, and malicious updates. For example, three hospitals may have different patient populations, so a strong aggregate score could hide poor results at one hospital. I would evaluate local performance as well as global quality and include the communication and privacy costs in the design.
Follow-up
Define whether the protected unit is a record or an entire client, and who may observe updates, before choosing a differential-privacy mechanism.