Apple uses the local differential privacy technique and the generation of synthetic data to train its system of artificial intelligence and improve functions such as Genmoji and writing tools, without putting user privacy at risk.
Apple ensures that you can collect information about users’ use trends without this implying a violation of their privacy, with the aim of improving your Apple Intelligence system. To do this, use a technique called local differential privacy.
With this technique, the technology company can “learn about the user community without learning about community individuals,” and the data identifiers are eliminated, which are sent to Apple channels encrypted.
The subsequent analysis of this data is withdrawn all information that can identify the user individually, such as IP and Metadata address and adds them to other users who have gone through the same process, to extract statistics from them that can be relevant.
These two stages “are carried out in a restricted access environment, so not even the privatized data are widely accessible to Apple employees,” says the company in the document that explains local differential privacy.
This technique is the one that Apple uses in some of the Apple Intelligence tools to improve them maintaining user privacy, although it requires that they have enabled the sharing of device data.
It is currently used in the Genmoji Emojis generator, to identify “popular indications and indications patterns”that the company ensures that it helps improve the response it generates, especially in requests that contain several entities, such as ‘dinosaur with cowboy hat’.
Apple has anticipated that the local differential privacy technique It will also be used in Image Playground, Image Wand, Memories Creation and writing tools in Apple Intelligence as well as in Visual Intelligence.
In the case of text functions that cannot take advantage of this technique, Apple uses synthetic data that “represent the trends added in real user data, without collecting emails or text of the devices.”
“Our goal is to generate synthetic electronic sentences or emails with a topic or style similar enough to the real to improve our summary models, but without Apple collects emails from the device,” he explains in his official blog. Likewise, it requires that users have enabled the sharing of device data.
The two techniques help “understand general trends, without the need to obtain personal information”, which allows Apple to improve Apple Intelligence functions.