Federated Learning is a decentralized approach to machine learning. It allows AI models to learn from data located on millions of devices (like your phone) without that raw data ever leaving the device.
Ready to simulate
The lifecycle of a single Federated Learning round.
The central server initializes a baseline machine learning model and sends this Global Model to participating edge devices.
Each device trains the model locally using its own private, raw data. The data never leaves the device.
Devices send back only what they've learned (model weights/gradients), not the data itself. This update is often encrypted.
The server aggregates the updates (e.g., via Federated Averaging) to create a smarter, new Global Model. The cycle repeats.
Raw personal data (photos, messages, health stats) remains on your device. The server only sees anonymous mathematical updates.
Because the model lives on your device, predictions happen instantly locally, without needing an internet connection to send data to the cloud.
While contributing to the global model, the local model on a specific device can be fine-tuned to uniquely fit that specific user's habits.
Data across devices isn't uniformly distributed (Independent and Identically Distributed). One user might only type in Spanish, another in English, skewing local models.
Devices have vastly different compute power, memory, battery levels, and network connectivity. Devices dropping out mid-training is a constant issue (Straggler Effect).
Malicious actors could spoof devices and send fake, manipulated model updates to deliberately ruin or backdoor the global model.