Differential Privacy
A mathematical guarantee that adding or removing one person's data from training barely changes the model.
Think of it like
It's like painting with someone in a crowd—whether one person is there or not barely affects the final picture.
Example
A model trained on loan applications uses differential privacy to ensure that whether one applicant's data is included or excluded, the model's predictions don't shift enough to expose them.
How it actually works
Differential privacy adds calibrated noise during training so no single person's data dominates the learning signal. You define a privacy budget (epsilon)—smaller epsilon means more noise and stronger privacy, but lower accuracy. It's a hard tradeoff: you get mathematical certainty that individual data can't be reverse-engineered, but performance often suffers.
For product teams
Offer privacy guarantees that regulators and users trust, even if it means slightly weaker model performance.
For engineers
Add Laplace or Gaussian noise to gradients during training, calibrated to your epsilon budget.
Related
- Privacy — Broader privacy approach.
- Federated Learning — Another privacy-preserving method.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome