If confidentiality and respect for private data is really not the strong point of artificial intelligence as a whole, Apple wants to play it differently. This is how the Cupertino company wants to operate differently from others.
During the sacrosanct WWDC 2024, Apple revealed its advances in artificial intelligence and in particular Apple Intelligence. An arsenal of innovative features integrating generative AI tools on its flagship devices: iPhone, iPad and Mac. Rewriting drafts, translating and rewriting texts, automatic correction or tailor-made creation of personalized emoticons, the features are quite numerous. Keen to reassure its customers on the issue of privacy, the company firmly insisted on the proven usefulness of these new technological trappings while guaranteeing the confidentiality of its users’ data. Should we believe it?
AI on two levels: local and cloud
Apple Intelligence is based on a hybrid approach, skillfully combining local and cloud processing capabilities. During its keynote, Apple explained how this organization will ensure the promptness and efficiency of usual tasks while rigorously preserving data confidentiality. Indeed, rudimentary artificial intelligence operations; transcription of calls or scheduling of diaries for example; will be processed directly within the device itself. However, for more complex queries involving personal data, the AI will request cloud servers to maintain a higher level of security.
Unlike other companies, Apple does not train its models on private data or user interactions. The training of these is based exclusively on licensed materials as well as publicly accessible online data collected by the Applebot web crawler. The company emphasizes that publishers can opt out of this collection and that it formally excludes sensitive data such as social security or credit card numbers, as well as certain inappropriate content.
Deep integration and optimization
One of the major advantages of Apple Intelligence lies in its tight integration into the company’s various operating systems and in-house applications. The artificial intelligence models are optimized to be energy efficient (a burning issue with AI) and adapted to the capabilities of iPhones (at least the most powerful ones, not everyone will have the right to this as you will see). This optimization allows AI queries to be processed locally, greatly reducing privacy concerns.
However, the other side of the coin remains in the use of local models of smaller size and necessarily less efficient, a bit like Stable Diffusion 3 Medium. To overcome this annoyance, Apple uses “fine-tuning” techniques which allow the precise adjustment of an existing artificial intelligence model for a specific task or domain. These specialized skills are then encapsulated in “adaptors” that can be changed as needed. Think of a general AI model as a Swiss army knife: it can perform many tasks, but not all with the same precision as a specialized tool. The “adapters” are like interchangeable blades, adding specific functionalities to the Swiss army knife, making it more efficient for particular tasks.
Apple claims to have optimized its models with the aim of speeding up the response time between sending a request and receiving the response. Several specific techniques therefore fully exploit the Neural Engine of Apple Silicon chips. Thus, only devices with M-series chips and the very recent iPhone 15 Pro and Pro Max will be able to use Apple Intelligence.
Enhanced security with Private Cloud Compute
When demands require more robust artificial intelligence models, Apple calls on its Private Cloud Compute (PCC) servers. These run on an iOS-based operating system by integrating a dedicated machine learning stack. The company highlights the impeccable security of its PCCs: secure boot and a Secure Enclave, a specialized processor for storing encryption keys, accessible only by the requesting device.
An end-to-end encrypted connection is established before any data transmission. Apple undoubtedly insists that it will not be able to access data within PCCs, as server management tools are absent. Furthermore, no data is stored persistently, ensuring that queries and personal data are erased after processing.
The hybrid approach of Apple Intelligence is certainly not completely new, with Google and Microsoft adopting similar strategies with mixed models. However, none of its competitors have placed as much emphasis on privacy. It now remains to be observed whether it will keep its promises, which are quite attractive on paper, it must be admitted. We’ll see this as Apple Intelligence rolls out throughout the year.
Source : The Verge
4