A machine learning approach where models are trained across multiple decentralized devices holding local data samples, without exchanging them, thereby preserving privacy.