Cross-Device Federated Learning is a federated learning setting involving millions of edge devices (smartphones, IoT sensors, equipment controllers) — each device has a tiny local dataset, limited compute, unreliable connectivity, and only a fraction participate in each training round.

Cross-Device Characteristics

• Many Participants: Millions to billions of devices (smartphones, sensors, controllers).
• Unreliable: Devices go offline, have intermittent connectivity, and varying compute capabilities.
• Tiny Local Data: Each device has very little local data — model must learn from many partial views.
• Asynchronous: No guarantee all selected devices complete their update within the time window.

Why It Matters

• Scale: Google trains keyboard prediction models on billions of phones using cross-device FL.
• Privacy at Scale: Each user's data stays on their device — no central data collection.
• Semiconductor IoT: Edge sensors in fabs could use cross-device FL for distributed monitoring models.

Cross-Device FL is learning from the edge swarm — training on millions of unreliable, resource-constrained devices for privacy-preserving intelligence at scale.