Skip to content
anythingai.space
AI Foundationsintermediate11 min · 9 steps

Privacy-Preserving ML: Federated Learning and Differential Privacy

Train on data that never leaves its owner, and add calibrated noise so no output reveals any one person.

raw dataTraining serverall raw dataPhone An = 100 examplesPhone Bn = 300 examplesPhone Cn = 600 examples
datarequestresponse
1 / 9
ProblemStep 1 of 9

Pooling everyone's raw data is the risky default

The usual recipe copies every user's data to one server and trains there. For keyboards, health records or bank transactions, that central pile is a breach target and often conflicts with regulations such as GDPR and HIPAA.

Hiding the raw data isn't enough on its own. A trained model or a published statistic can still reveal individuals, for example through membership inference (was this person in the training set?) or by subtracting two aggregate answers.

Watch outTwo separate risks: where the data goes, and what the outputs reveal. Federated learning addresses the first, differential privacy the second.

Lesson 34 of 34 in AI Foundations