How Apple trains its AI models using synthetic data & differential privacy

In a blog article, Apple outlined its strategy for training its artificial intelligence (AI) models that prioritise user privacy and avoid the direct collection or copying of personal content from users' iPhones or Macs. Instead, the company will primarily utilise synthetic data and differential privacy techniques.

The article explains how this works for a range of features, from improving Genmoji to text generation.  

For text summarisation and writing tools used in longer format content, Apple's AI models compare synthetic, email-like messages against a small, anonymised sample of a real user's content. This comparison will occur locally on the user's device. The device will then identify which synthetic messages most closely resemble its user sample and send aggregated information about this match back to Apple. The actual user data remains on the device and is never transmitted to Apple.

Such privacy-preserving techniques allow Apple to improve its AI models for tasks involving longer-form text generation without needing to access or store private user content. The strategy builds upon Apple's existing use of differential privacy, a method that introduces random noise into datasets to protect individual identities. Apple has employed differential privacy since 2016 to understand usage patterns while adhering to its strong privacy safeguards.