Question

A mobile keyboard app wants to improve its next-word prediction model by learning from the words users type, but the privacy engineering team insists that no individual user's raw typed text should ever leave the device or be reconstructable by the company's central servers. The team wants a technique that adds privacy protection at the point of collection on each device, before any data is transmitted. Which approach best meets this requirement?

Accepted Answer

Apply local differential privacy by adding calibrated noise to each device's contribution before it is sent to the server

Answer

Collect the raw typed text over TLS and apply central differential privacy noise to aggregate query results at the server

Answer

Pseudonymize each user's typed text with a per-user token before transmission and store the mapping in a separate vault

Answer

Hash the typed text with a salted SHA-256 function on the device before sending it to the server for aggregation