Privacy-preserving AI brings together methods that enable training and using machine learning models without directly sharing raw data. Core technologies include Federated Learning (FL), Differential Privacy (DP), secure computation (homomorphic encryption and secure multi‑party computation), and hardware trusted execution environments. This article systematically reviews FL architectures and algorithms (FedAvg, FedProx, SCAFFOLD, personalization), DP mechanisms (central and local models, DP‑SGD, PATE, privacy budget accounting), and secure aggregation protocols along with defenses against poisoning and backdoors. We discuss leakage metrics (membership/property inference AUC, advantage), engineering trade‑offs among utility, privacy, and cost, and regulatory requirements (privacy by design, DPIA).