Federated learning trains models on decentralized data without centralizing raw data, preserving privacy. Mechanism: Central server sends model to devices/clients, each client trains on local data, clients send model updates (not data) to server, server aggregates updates (FedAvg: average weights), repeat until convergence. Privacy benefits: Raw data never leaves device, only model updates transmitted, can combine with differential privacy on updates. Applications: Mobile keyboards (next word prediction), healthcare (cross-hospital learning), finance (fraud detection across banks), IoT devices. Challenges: Non-IID data: Client data differently distributed, hurts convergence. Communication: Model updates expensive to transmit frequently. Device heterogeneity: Different compute capabilities. Stragglers: Slow clients delay rounds. Adversarial clients: May send malicious updates. Aggregation methods: FedAvg (weighted average), FedProx (regularization), personalized variants. Privacy considerations: Updates can still leak information - use secure aggregation, differential privacy. Frameworks: TensorFlow Federated, PySyft, Flower. Trade-offs: Privacy vs accuracy vs communication cost. Enables ML where data sharing is impossible.