What is cross-device federated learning?

Cross-device federated learning (FL) is a decentralized machine learning approach where models are trained collaboratively across a large number of devices—such as smartphones, IoT sensors, or edge devices—without centralizing raw data. Unlike traditional methods that require data to be uploaded to a server, FL keeps data on local devices. Each device trains a shared model using its own data and sends only model updates (e.g., gradients or parameters) to a central server. These updates are aggregated to improve the global model, which is then redistributed to devices for further training. This process repeats iteratively, enabling the model to learn from diverse data sources while preserving user privacy.

A typical cross-device FL workflow involves three steps. First, a central server initializes a global model and selects a subset of available devices (e.g., thousands of phones) to participate in a training round. Each device downloads the model, trains it locally using its data, and computes updates. For example, a smartphone keyboard app might train a next-word prediction model using the user’s typing history. After local training, devices send encrypted updates back to the server. The server then aggregates these updates—often using methods like Federated Averaging—to create a new global model. This cycle continues until the model converges. Key technical considerations include handling intermittent device availability (since devices may disconnect), optimizing communication efficiency (to reduce bandwidth), and ensuring compatibility across heterogeneous hardware.

Implementing cross-device FL poses challenges. Device heterogeneity—varying computational power, storage, and network conditions—can lead to uneven training times or dropouts. For instance, a low-end phone might take longer to train a model than a high-end device, delaying aggregation. Privacy remains critical; even though raw data isn’t shared, updates could leak sensitive information. Techniques like secure aggregation (masking updates with encryption) and differential privacy (adding noise to updates) mitigate these risks. Frameworks like TensorFlow Federated or PyTorch’s FL libraries provide tools to manage device coordination, model distribution, and aggregation. A real-world example is training a health-monitoring model across smartwatches without exposing users’ heart rate or activity data. By addressing these challenges, cross-device FL enables scalable, privacy-aware ML applications in scenarios where data cannot be centralized.