Federated Learning
Train a shared model across many devices or institutions — without any raw data ever leaving its source.
Federated Learning (FL) inverts the usual machine learning workflow: instead of moving data to a central server to train a model, FL moves the model to the data. Each participant — a bank, a hospital, a mobile device — trains a copy of the model locally on its own data, then shares only the resulting model updates, never the underlying records.
A central or distributed aggregator combines those updates into a single improved global model, which is redistributed for the next round of local training. The raw data behind each contribution never leaves its source, which is what lets banks, telcos, and hospitals collaborate on shared models across institutional and regulatory boundaries that would otherwise block pooling data outright. PryvX pairs FL with secure aggregation and differential privacy so even the model updates themselves can’t be reverse-engineered into individual records.
How It Works
Federated Learning in four steps
Train Locally
Each device or institution trains the model on its own local data — the raw data never leaves its source.
Share Updates
Only model updates (gradients or weights), not raw data, are sent to a central or distributed aggregator.
Aggregate
Updates from all participants are combined into a single improved global model.
Distribute
The updated global model is sent back to every participant, and the cycle repeats.
Real-World Applications
Where FL shows up in production
Cross-Bank Fraud & Credit Models
Banks and telcos jointly improve fraud-detection or credit-risk models without centralizing customer records — see our Finance use case.
On-Device Personalization
Predictive text, recommendation, and personalization models improve from user behavior without raw usage data ever leaving the device.
Multi-Hospital Clinical Research
Hospitals train diagnostic models across patient populations without any institution exposing its raw patient records to the others.
Telecom Network Optimization
Operators improve network and fraud models using signals from subscriber devices and towers without centralizing raw telemetry.