Federated Learning Quiz

Answers and explanations appear as you take the quiz.

1. How does federated learning train a shared model without centralizing user data?
   • Train one model by uploading everyone's raw data to a central server
   • Use only synthetic data and never real data
   • Train a shared model by sending model updates instead of raw data to a server
   • Run a separate, unrelated model on each device forever
2. What does the Federated Averaging (FedAvg) algorithm do on the server?
   • Deletes all client models
   • Averages the weight updates from clients, usually weighted by data amount
   • Sends raw data back to clients
   • Trains the model entirely by itself
3. Which product famously used federated learning to improve predictions?
   • A central spam email filter
   • A weather satellite
   • A video game console
   • Google's Gboard keyboard
4. Why might a hospital collaboration investigate federated training?
   • It can train through local computations without centrally pooling patient records
   • It automatically makes all sharing legally compliant
   • It removes every risk of revealing patient information
   • It requires no evaluation of the resulting model
5. Why are model updates alone not automatically fully private?
   • Updates are always random noise
   • Updates can still leak information about the underlying data
   • Updates contain the raw data unchanged
   • Updates cannot be sent over networks
6. What problem does 'non-IID' data cause in federated learning?
   • All devices have identical data, so nothing is learned
   • The server runs out of memory instantly
   • The model can only learn images
   • Participants' data distributions differ a lot, making averaging harder
7. Two FedAvg clients return scalar weights 0 and 10 after using 1 and 3 training examples. What is the data-weighted average?
   • 7.5
   • 5
   • 10
   • 2.5
8. What does a successful secure-aggregation protocol aim to reveal to the federated server?
   • Every client’s raw training record
   • Only a participant’s device password
   • The aggregate update rather than each participant’s individual value
   • A complete guarantee against all information in the aggregate