A model trained on data nobody can see will still give some of it back. Membership inference, model inversion and gradient reconstruction have each turned a shipped model into a disclosure, and regulators have begun writing the mathematics of leakage into enforceable law. Privacy is now an engineering property of the system, and it belongs to the people who build it.

This book treats the four families of privacy-enhancing technology, synthetic data, federated learning, differential privacy, and the cryptographic and hardware methods, as complementary rather than competing. Each is introduced through a concrete problem before it is formalized, every term is defined where it first appears, and the worked examples reward a reader with a pencil.

Written for data scientists, platform engineers, architects and the technical leaders who answer for what their systems do with personal data. It assumes you have shipped machine learning systems, and no background at all in privacy theory, cryptography or data-protection law.