Cross-Device Federated Learning

The defining characteristic is scale, often involving millions to billions of edge devices such as smartphones, wearables, or IoT sensors. This vast population creates a highly statistically diverse but operationally chaotic training environment.

• A single training round may sample thousands of devices.
• The server never communicates with the entire population simultaneously.
• Population size provides inherent statistical strength if harnessed correctly.

Clients are ephemeral participants. They are only available for computation when meeting specific local criteria: typically, the device is idle, connected to an unmetered Wi-Fi network, and charging. A client can drop out of a training round at any moment without warning.

• Eligibility checks gate participation.
• The system must be robust to a high dropout rate (often >90%).
• Synchronous coordination is impossible; protocols are inherently asynchronous.

Unlike server-grade hardware in cross-silo FL, edge devices possess minimal compute, memory, and battery. Model architectures and update computations must be aggressively optimized.

• On-device memory limits model size (often <100MB).
• Battery impact must be imperceptible to the user.
• Training is restricted to spare compute cycles.
• This necessitates techniques like model compression and quantization before local training.

Data generation is intrinsically tied to user behavior, creating extreme statistical heterogeneity. A single device's local dataset reflects a unique, narrow slice of the world.

• Pathological non-IIDness: Data distributions vary wildly in both features and labels.
• Quantity skew: Some users generate gigabytes of data; others generate kilobytes.
• This violates the IID assumption of standard optimization, requiring algorithms like FedProx or SCAFFOLD to correct for client drift.

The raw data on a device is the ultimate personal identifier. Cross-device FL is predicated on the absolute guarantee that raw data never leaves the device. Privacy is not an add-on; it is the architectural premise.

• Only encrypted model updates (gradients or weights) are transmitted.
• Augmented with Secure Aggregation to prevent the server from inspecting individual updates.
• Often combined with Local Differential Privacy to provide a formal mathematical guarantee against membership inference.

Uploading a model update over a cellular network is often slower and more energy-intensive than the local training computation itself. Communication efficiency is the paramount optimization target.

• Techniques like gradient compression (sparsification, quantization) reduce payload size.
• Federated Averaging minimizes communication rounds by performing multiple local epochs.
• Protocols are designed to be bandwidth-adaptive and resilient to slow links.