Meta 已将“私有处理”（Private Processing）功能扩展至 AI 眼镜，利用机密虚拟机（CVM）处理敏感 AI 工作负载，并在云端持久化保存用户的上下文信息。该设计融合了可信执行环境（TEE）、远程认证、匿名凭证、OHTTP 以及加密存储技术。Meta 表示，该功能支持流式转录、上下文搜索以及需要更多计算能力和持久状态的长期记忆召回功能。
通话和消息等基础交互可以在 AI 眼镜上运行，但更复杂的工作负载则需要更大的模型，并且需要访问随时间积累的个人上下文信息。Meta 于 2025 年针对 WhatsApp 和 Meta AI 推出了“私有处理（Private Processing）”功能，现在将其扩展到了可穿戴 AI 工作负载。
在 LinkedIn 上，Meta 公司 XR 技术副总裁 Oskar Linde 将这一方法描述为：在确保 Meta 无法访问相关数据的同时，让 AI 能在全天使用中发挥更大的作用，而且满足 AI 眼镜所需的有状态、个性化场景需求。
该架构将机密计算边界从设备端转移至 Meta 的云基础设施中。AI 模型在由硬件 TEE 支持的机密虚拟机（CVM）内运行，其中配备了处理器加密的内存。Meta 表示，其边界覆盖 CPU 和 GPU，可以确保混合工作负载始终受到保护。
在传输个人上下文之前，AI 眼镜会执行远程认证，并与 TEE 建立经过认证的 TLS 会话。设备会验证经硬件签名的证书，并将工作负载测量值与仅可追加的透明账本进行比对。如果验证失败，则将阻止数据传输。匿名凭证以及由 Fastly 或 Cloudflare 运营的第三方 OHTTP 中继则进一步将用户身份与请求处理分离。
持久化状态始终保留在机密计算边界之内。Meta 将存储引擎置于 TEE 内，并在数据离开受保护环境之前使用用户提供的密钥对其进行加密。查询也是在同一边界内执行，限制了数据库访问模式的暴露。
私有处理基于硬件隔离、故障关闭保证、公开可验证性、不可被针对性识别以及加密存储构建。生产环境 CVM 镜像会在透明账本中进行注册。由于工程师无法检查受保护的内存，也无法捕获模型的输入和输出，所以 Meta 使用了 CPU 利用率、内存分配、网络延迟和硬件故障率等聚合运营信号。
始终在线的可穿戴 AI 所涉及的隐私问题也引发了社区讨论。在 Reddit 上一个关于 Meta 眼镜的讨论中，有评论者提出，要让智能助手提供上下文相关信息，就必须持续监听、观测周边环境，这也凸显了持续感知模式下的隐私风险问题。
部署方面仍然存在局限。WIRED 杂志报道称，Meta AI 眼镜目前尚未支持“私有处理”功能，Meta 也未明确说明哪些功能将率先采用该系统，也未提供具体的上线时间表。Meta 工程总监 Pritam Shah 表示，该平台已经准备就绪而且运行正常，但 Meta 在广泛部署前进行公告，是为了与安全和隐私研究人员以及第三方审计机构展开合作。
Meta 表示，正与包括 NCC Group 在内的独立安全公司及安全研究人员合作，共同审查该系统的架构和认证机制。随着该系统为支持具备状态感知能力、多模态及自主性的 AI 体验而不断扩展，该公司还计划向研究人员提供 CVM 二进制文件、工具及文档。
原文链接：https://www.infoq.com/news/2026/10/meta-private-process-ai-glasses/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Meta has extended the "Private Processing" function to AI glasses, using the CVM to handle sensitive AI task loads and to keep user context information in clouds for a lasting period. The design incorporates a credible implementation environment (TEE), remote authentication, anonymous vouchers, OHTTTP and encryption storage techniques. Meta states that this function supports streaming, context searches and long-term memory recall functions that require additional computing capabilities and durability.
Intersections such as calls and messages can run on AI glasses, but more complex loads require larger models and access to personal context information accumulated over time. Meta introduced the "Private Processing" feature in 2025 for whatsApp and Meta AI, and now extends it to the wearable AI workload.
On LinkedIn, Meta XR Technical Vice President Oskar Linde describes this approach as a way for AI to play a greater role in full-time use while ensuring that Meta does not have access to relevant data, and to meet the state and personalized scene needs of AI glasses.
The structure transferred the confidential calculation boundary from the end of the equipment to the cloud infrastructure in Meta. The AI model operates in a confidential virtual machine (CVM) supported by the hardware TEE, with an encrypted memory of the processor. Meta states that its borders cover CPU and GPU and ensure that mixed loads are always protected.
Before transmitting a personal context, AI glasses perform remote authentication and create a certified TLS session with TEE. The equipment validates the hardware-signed certificate and compares the task load measurement with only additional transparent books. If the validation fails, the data transfer is stopped. Anonymized certificates and third party OHTTP relays operated by Fastly or Cloudflare further separate user identity from request processing.
Endurance remains within the confidential calculation boundary. Meta places the storage engine in TEE and encrypts the data with the key provided by the user before leaving the protected environment. Queries are also carried out within the same border, limiting the exposure of the database access mode.
Private processing is based on hardware isolation, fail-off assurance, open authentication, non-target identification and encrypted storage construction. CVM mirrors are registered in transparent books. As engineers were unable to check protected memory or capture model input and output, Meta used such aggregate operating signals as CPU utilization, memory distribution, network delay and hardware failure rates.
The privacy implications of the continuous online wearable AI have also triggered community discussions. In a previous discussion on Meta glasses in Reddit, some commentators suggested that, in order for smart assistants to provide information on context, continuous listening and observation of the surrounding environment was necessary, which also highlighted privacy risks in the context of a continuous perception model.
Deployment constraints remain. WIRED magazine reports that Meta AI glasses do not currently support the “private handling” function, nor does Meta specify which functions will be the first to adopt the system, nor does it provide a specific go-live schedule. Meta Engineering Director Pritam Shah stated that the platform was ready and functioning, but that Meta had been advertised before extensive deployment to work with security and privacy researchers and third-party audit institutions.
Meta indicated that the architecture and certification mechanisms of the system were being reviewed in cooperation with independent security companies, including NCC Group, and security researchers. As the system expands to support the AI experience of state perception, multimodules and autonomy, the company also plans to provide CVM binary files, tools and documents to researchers.
Original link: https://www.infoq.com/news/2026/10/meta-private-process-ai-glasses/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.