Federated learning and differential privacy

In short: Federated learning trains a model on many devices without the raw data leaving the device. Differential privacy adds noise so that individuals cannot be reconstructed from the model.

In more detail: Example: a phone keyboard learns word suggestions from typing behaviour without texts travelling to the manufacturer.

In detail

• Process: the server sends the model → devices train locally → send only weight changes → the server averages (federated averaging).
• Problems: different data on each device, slow connections, weight changes can still reveal information.
• Differential privacy: every change is clipped and noise is added. A mathematical guarantee (ε) that a single record has hardly any influence.
• Secure aggregation: encrypted sums so the server sees only the average.
• Edge AI: models run directly on the phone or sensor without the cloud: smaller, faster, more private.
• Alternative: anonymise data (hard: re-identification) or use synthetic data.
• Use: keyboards, health data between hospitals, traffic data.

See also: Training data, copyright and privacy in AI, Model compression: distillation and pruning