Federated learning trains a shared model across many devices without ever centralizing the raw underlying data — a genuinely clever privacy-preserving approach, notably used by Google for features like predictive keyboard text.
1. A central server sends the current model to many participating devices 2. Each device trains the model briefly on ITS OWN LOCAL data (the raw data NEVER leaves the device) 3. Devices send back only the MODEL UPDATES (gradients/weight changes), not the actual raw data itself 4. The server aggregates these updates from many devices into an improved shared model 5. Repeat - the shared model improves over many rounds, without the server ever seeing anyone's individual raw data directly
Why This Is Genuinely Different from Normal Training
Traditional model training centralizes data on one server or cluster before training against it directly. Federated learning inverts this specific pattern — the model travels to where the data already lives (each individual device), trains briefly and locally, and only the resulting learned updates travel back, not the raw, potentially sensitive underlying data itself.
The Real Trade-offs
Federated learning is genuinely more complex to implement and coordinate than centralized training (handling unreliable device connectivity, non-uniform data distributions across devices, and aggregating updates correctly and securely), and it isn’t a complete privacy panacea — model updates can theoretically leak some information about the underlying local training data under certain sophisticated attacks. It’s a genuine, meaningful improvement over full data centralization, not an absolute, mathematically perfect privacy guarantee.