Chapter 17 · AI and society
Privacy-preserving AI
Federated learning trains on devices without collecting data; differential privacy adds noise so no one person stands out.
Chapter 17 · AI and society
Federated learning trains on devices without collecting data; differential privacy adds noise so no one person stands out.
Photos and messages stay on each phone…
In federated learning the training goes to the data instead of the data going to the server. Only small updates travel.
Normally, to train a model on what people type or photograph, a company gathers all of it in one place. That pile can be leaked, misused or demanded by others.
Each dot is hours of sleep on a line from 1 to 11. Your phone adds noise before sending. The server only sees the pink dot, yet it wants the group average.
your real valuewhat you sendaverage the server works out
One person’s dot is blurred, yet the group average stays close. Crowds make privacy affordable. A simplified toy with invented data; real systems use carefully proven methods.
What leaves the phone in federated learning?
Why does noise barely hurt the group average in a large crowd?
What does turning the noise up do?