Apple Unveils Privacy-Savvy On-Device AI Training with User Data

Apple's New On-Device AI Training System

Apple has introduced a new on-device training system for its artificial intelligence models, which leverages user data while maintaining privacy. The company had previously delayed the rollout of advanced Siri features and is now shifting its approach to improve these capabilities.

Synthetic Data Limitations

Apple traditionally uses synthetic data to train AI models, but this method has limitations in capturing complex trends, especially for tasks like summarization or writing tools that require understanding longer sentences or email messages.

New Privacy-Preserving Technology

To enhance model training, Apple is implementing a technology that compares synthetic emails with a small sample of recent user emails. This process involves:

• Synthetic Message Generation: Creating a large set of synthetic messages on various topics.
• Embedding Creation: Generating embeddings for these synthetic messages to capture key dimensions like language, topic, and length.
• Device Analytics: Sending these embeddings to a small number of user devices that have opted into Device Analytics.
• User Email Embeddings: Participating devices compute embeddings for a small sample of recent user emails and determine the closest synthetic embeddings.
• Differential Privacy: Apple uses differential privacy techniques to learn the most frequently selected synthetic embeddings without identifying individual users.
• Data Refinement: The top synthetic embeddings are used to generate or refine training data, improving topics and language in tasks like email summaries.

Rollout Plans

Apple plans to implement this system in a future beta of iOS 18.5 and macOS 15.5. The company asserts that these techniques allow it to understand overall trends without compromising individual user privacy.