A mobile OS company wants to improve its keyboard autocomplete model using the typing behavior of 200 million active users. A privacy engineer objects to collecting all keystrokes to a central server, noting the sensitivity of typed data (messages, passwords, health queries). She proposes an alternative training architecture that keeps the data on users' phones.
A data scientist proposes using federated learning for training a mobile keyboard autocomplete model. What is the primary privacy benefit of federated learning compared to centralized training?
- A.A. Federated learning encrypts model weights so they cannot be inspected by users
- B.D. Federated learning trains on synthetic data, so real user data is never involved
- C.C. Federated learning uses differential privacy automatically, providing mathematical privacy guarantees
- D.B. Federated learning keeps raw user data on-device; only model updates (gradients) are sent to the server, reducing data centralization
Why D is correct
In federated learning, each device trains the model locally on its data and sends only gradient updates (not raw data) to a central server for aggregation. This avoids centralizing sensitive data - critical for keyboard inputs which may contain passwords, medical searches, and personal messages. Federated learning does not inherently encrypt weights for users, does not automatically apply DP (though they are often combined), and does not use synthetic data.
Know someone studying for AI Security Fundamentals? Send them this one.