Tomcat
Professional
- Messages
- 2,689
- Reaction score
- 916
- Points
- 113
How does the company protect our data and ensure our privacy?
Apple has introduced Apple Intelligence, a set of features that bring generative AI capabilities, such as rewriting draft emails, summarizing notifications, and creating custom emojis, to the iPhone, iPad, and Mac. During the presentation at the WWDC conference, representatives of the company spoke in detail about the benefits of the tools themselves. Much attention has also been paid to the promise of complete privacy when using new AI tools.
According to the developers, a high level of privacy is provided by Apple's two-way approach. On the one hand, Apple Intelligence works locally on the device, which allows you to quickly solve basic AI tasks. However, for more complex requests that require the transfer of personal data, cloud servers are used.
The key feature is Apple's use of its own AI models. Unlike competitors, the company does not train them on private data or information about user interaction. Instead, it uses licensed materials and public information collected by the Applebot crawler. Authors can prohibit indexing their content, just like Google and OpenAI. Credit cards, social security numbers, and obscene language are also excluded from training data.
One of the main advantages of Apple Intelligence is its deep integration into Apple operating systems and applications, as well as optimizing models for energy efficiency and size to work on the iPhone. Local processing of requests removes many concerns about privacy, but at the same time it is necessary to use more compact and limited AI tools.
To improve the effectiveness of local models, Apple uses the method of final tuning (fine-tuning), specifically training them for specific tasks such as spell checking or text summarization. These specialized skills are implemented as "adapters" that can be flexibly connected to the basic model for a specific task, giving it new features.
To speed up the work of AI, the company uses special techniques such as "speculative decoding"," selective context processing "and"grouping similar queries". For this purpose, the neural cores of Apple Silicon processors are used. Chipmakers have recently begun to integrate specialized neuroprocessors into new systems on a chip, which allows you to offload the CPU and GPU from processing machine learning and AI algorithms. That's why Apple Intelligence only works on devices with M-series chips, including the iPhone 15 Pro and Pro Max.
According to Apple's internal research, out of 750 analyzed text summarization responses, the company's local AI model (with the appropriate adapter) showed results that were more attractive to people than the Microsoft Phi-3-mini model. This sounds like a big achievement, but modern chatbots usually use much more powerful cloud models to get better results. Here, Apple tries to find a balance between quality and privacy, offering seamless sending of complex requests to cloud servers with confidential data processing.
If a request requires a more productive AI model, Apple sends it to its Private Cloud Compute (PCC) servers. PCC runs on its own operating system (based on iOS) and has its own software stack for Apple Intelligence. According to the company, the PCC has its own Secure Enclave module for storing encryption keys that are compatible only with the requesting device. Also, a special monitor ensures that only verified code is running on the PCC.
Before sending the request, the user's device creates a connection to the PCC cluster that is protected with end-to-end encryption. Apple claims that it can't access data on the PCC because the servers lack remote management tools and a command shell. There is also no persistent storage on the PCC, so requests and any personal data from the Apple Intelligence semantic index are deleted after processing in the cloud.
Each PCC build will have a public virtual version for research audit. Only signed and registered builds that have passed validation will be implemented in the production environment.
Another way that Apple is coping with privacy concerns is by shifting this issue to third-party companies. The updated Siri voice assistant can redirect some complex queries to OpenAI's ChatGPT cloud, but only with the user's permission after they ask a really difficult question. According to Apple CEO Tim Cook in an interview with Marquez Brownlee, the ChatGPT system will be connected to answer general questions that go beyond the personal context.
Apple isn't the first company to combine on-premise and cloud-based data processing for its AI tools. Google uses the on-premises Gemini Nano model on Android devices, along with the cloud-based Pro and Flash models. Microsoft is also leveraging local processing on Copilot Plus computers, drawing on OpenAI resources and developing its own MAI-1 model. However, none of Apple's competitors has yet placed such a strong emphasis on the obligation to preserve the confidentiality of user data.
Of course, all this looks impressive in prepared demonstrations and official documents. But right now, the most important thing for researchers is to make sure that Apple Intelligence is effective in practice when it becomes available later this year.
Apple has introduced Apple Intelligence, a set of features that bring generative AI capabilities, such as rewriting draft emails, summarizing notifications, and creating custom emojis, to the iPhone, iPad, and Mac. During the presentation at the WWDC conference, representatives of the company spoke in detail about the benefits of the tools themselves. Much attention has also been paid to the promise of complete privacy when using new AI tools.
According to the developers, a high level of privacy is provided by Apple's two-way approach. On the one hand, Apple Intelligence works locally on the device, which allows you to quickly solve basic AI tasks. However, for more complex requests that require the transfer of personal data, cloud servers are used.
The key feature is Apple's use of its own AI models. Unlike competitors, the company does not train them on private data or information about user interaction. Instead, it uses licensed materials and public information collected by the Applebot crawler. Authors can prohibit indexing their content, just like Google and OpenAI. Credit cards, social security numbers, and obscene language are also excluded from training data.
One of the main advantages of Apple Intelligence is its deep integration into Apple operating systems and applications, as well as optimizing models for energy efficiency and size to work on the iPhone. Local processing of requests removes many concerns about privacy, but at the same time it is necessary to use more compact and limited AI tools.
To improve the effectiveness of local models, Apple uses the method of final tuning (fine-tuning), specifically training them for specific tasks such as spell checking or text summarization. These specialized skills are implemented as "adapters" that can be flexibly connected to the basic model for a specific task, giving it new features.
To speed up the work of AI, the company uses special techniques such as "speculative decoding"," selective context processing "and"grouping similar queries". For this purpose, the neural cores of Apple Silicon processors are used. Chipmakers have recently begun to integrate specialized neuroprocessors into new systems on a chip, which allows you to offload the CPU and GPU from processing machine learning and AI algorithms. That's why Apple Intelligence only works on devices with M-series chips, including the iPhone 15 Pro and Pro Max.
According to Apple's internal research, out of 750 analyzed text summarization responses, the company's local AI model (with the appropriate adapter) showed results that were more attractive to people than the Microsoft Phi-3-mini model. This sounds like a big achievement, but modern chatbots usually use much more powerful cloud models to get better results. Here, Apple tries to find a balance between quality and privacy, offering seamless sending of complex requests to cloud servers with confidential data processing.
If a request requires a more productive AI model, Apple sends it to its Private Cloud Compute (PCC) servers. PCC runs on its own operating system (based on iOS) and has its own software stack for Apple Intelligence. According to the company, the PCC has its own Secure Enclave module for storing encryption keys that are compatible only with the requesting device. Also, a special monitor ensures that only verified code is running on the PCC.
Before sending the request, the user's device creates a connection to the PCC cluster that is protected with end-to-end encryption. Apple claims that it can't access data on the PCC because the servers lack remote management tools and a command shell. There is also no persistent storage on the PCC, so requests and any personal data from the Apple Intelligence semantic index are deleted after processing in the cloud.
Each PCC build will have a public virtual version for research audit. Only signed and registered builds that have passed validation will be implemented in the production environment.
Another way that Apple is coping with privacy concerns is by shifting this issue to third-party companies. The updated Siri voice assistant can redirect some complex queries to OpenAI's ChatGPT cloud, but only with the user's permission after they ask a really difficult question. According to Apple CEO Tim Cook in an interview with Marquez Brownlee, the ChatGPT system will be connected to answer general questions that go beyond the personal context.
Apple isn't the first company to combine on-premise and cloud-based data processing for its AI tools. Google uses the on-premises Gemini Nano model on Android devices, along with the cloud-based Pro and Flash models. Microsoft is also leveraging local processing on Copilot Plus computers, drawing on OpenAI resources and developing its own MAI-1 model. However, none of Apple's competitors has yet placed such a strong emphasis on the obligation to preserve the confidentiality of user data.
Of course, all this looks impressive in prepared demonstrations and official documents. But right now, the most important thing for researchers is to make sure that Apple Intelligence is effective in practice when it becomes available later this year.
