Apple to Train AI with On-Device Data—No Emails or Personal Content Needed
3 days ago | 5 Views
Apple is unveiling a novel strategy for training its artificial intelligence (AI) models, designed to improve their effectiveness while safeguarding user privacy. This new initiative is expected to be implemented in the forthcoming beta versions of iOS 18.5 and macOS 15.5, as reported by Bloomberg. Additionally, Apple has shared insights on this development through a blog post on its Machine Learning Research website.
At present, Apple utilizes synthetic data—information generated artificially rather than sourced from real users—to train various AI functionalities, including writing assistants and email summaries. Although this method protects user privacy, Apple recognizes that synthetic data has its limitations, particularly in accurately reflecting trends in how individuals compose or summarize longer messages.
New Privacy-Focused Training Method
To overcome this limitation, Apple has created a technique that compares synthetic data with actual user data without ever accessing the content of user emails. The process is as follows: Apple produces thousands of fictitious emails that cover a range of common topics, such as a request to play tennis at a specific time. Each email is transformed into an embedding, which is a data representation that encapsulates its content, including its subject and length.
These embeddings are transmitted to a select group of devices whose users have opted into Apple's Device Analytics program. The devices then compare the synthetic embeddings with a sample of the user's recent emails, identifying the most similar synthetic message. Importantly, Apple guarantees that the actual emails and their corresponding results remain solely on the device.
Differential Privacy Safeguards Data Security
The company employs a privacy method known as differential privacy, which ensures that only anonymous signals are sent back from the device. Apple analyzes the frequency with which specific synthetic messages are chosen without linking the selections to individual devices. This anonymous data aids in enhancing AI features, making them more reflective of the types of content users typically compose, all while safeguarding user privacy.
Apple has previously implemented this method in other features, such as Genmoji, its custom emoji creation tool. In this instance, Apple monitors which emoji prompts gain popularity, such as “an elephant in a chef's hat,” to refine its AI model. However, uncommon or unique prompts are excluded from the analysis, ensuring that user-specific data is never associated with any particular device.
Looking ahead, Apple intends to broaden this privacy-centric approach to additional AI tools, including features like Image Playground, Image Wand, and Visual Intelligence, while consistently prioritizing user privacy.
Get the latest Bollywood entertainment news, trending celebrity news, latest celebrity news, new movie reviews, latest entertainment news, latest Bollywood news, and Bollywood celebrity fashion & style updates!
