Photo credit: www.theverge.com
Apple Innovates AI Model Training While Protecting User Privacy
In a significant step towards enhancing its artificial intelligence capabilities, Apple has announced a new method to improve its AI models without the need to access or replicate data from users’ devices. This approach was detailed in a recent blog post that highlighted how devices will utilize a synthetic dataset to assess real-time user interactions, specifically focusing on recent emails and messages from individuals who have opted into Apple’s Device Analytics program.
The technology enables Apple devices to compare synthetic data with actual user samples, determining which synthetic inputs most closely resemble the real data. The critical aspect of this system is that only a signal indicating the best match is sent back to Apple, ensuring that no personal data leaves the device. By identifying the most frequently selected synthetic samples, Apple plans to refine features like email summaries and other AI text outputs.
Currently, Apple’s AI systems rely solely on synthetic data, which some critics argue may lead to less effective results. The challenges faced by Apple in rolling out its premier AI features have been notable, prompting delays in capabilities and changes in leadership within the Siri development team, as noted by Bloomberg‘s Mark Gurman.
In light of these challenges, Apple is making strides to revamp its approach through a beta version of its operating systems—iOS and iPadOS 18.5, along with macOS 15.5—as it introduces its innovative AI training system.
Since the introduction of iOS 10 in 2016, Apple has emphasized its commitment to user privacy, utilizing techniques like differential privacy to safeguard personal information. This approach has been successful in enhancing features like Genmoji, and the same privacy principles will underpin the new AI training strategy, ensuring that the incorporation of randomized information into larger datasets protects individual identities from being identifiable.
Source
www.theverge.com