This chapter examines how federated learning (FL) systems can be designed to protect user privacy. It begins by discussing the risks of privacy leakage when model parameters are shared across devices, even without transmitting raw data. The chapter then introduces formal measures to quantify privacy, including sensitivity analysis and differential privacy (DP). It presents two main approaches to enforcing DP in FL: adding random noise to model updates (postprocessing) and modifying data before training (preprocessing). The chapter also introduces Rényi differential privacy as a more flexible alternative to standard DP. In addition, it explores private feature learning, which involves transforming features to reduce the risk of exposing sensitive attributes while preserving task performance. The chapter concludes with practical strategies, including the privacy funnel method and optimal linear transformations, to balance utility and privacy. These tools provide a foundation for building FL systems that comply with data protection regulations while maintaining high model quality.