869 points • 5 days ago • Article Link

自 macOS 13 起，Apple 已悄然内置了一个复杂的后台系统，该系统由两个主要守护进程（daemons）组成，其共同目的是构建用户的详细个人知识图谱。该系统并非孤立运行，而是会汇总 Mac 上来自多种来源的数据，包括通讯录（Contacts）、日历（Calendar）、照片（Photos）、消息（Messages）、邮件（Mail）、位置历史记录以及屏幕使用时长（Screen Time）。通过交叉比对这些输入，系统生成一个可查询且持久化的数据库，用于追踪你的身份信息、家庭与社交关系、行为模式，甚至你的身体活动与日常作息。

这一机制的核心位于目录 ~/Library/IntelligencePlatform，其中包含 51 个 SQLite 数据库。最关键的是 graph.db，这是一个三元组存储（triple store），用 181 个实体类和 278 个谓词来组织信息。该图谱记录实体之间的关系，并为每条连接附上置信度分数，同时注明该推断的来源。尽管系统也会采用较直接的方法，比如从联系人名片或日历条目复制数据，但它还使用了更先进的技术，包括基于 CoreML 的 XGBoost 分类器、嵌入模型以及本地 LLM，用于对关系进行分类、并识别个人身份。

该系统特别敏感的一点在于，它能够在没有明确用户输入的情况下推断亲属关系与社交亲密程度。例如，Apple 预装了 XGBoost 分类器，通过分析 144 个不同的行为特征（如消息发送频率、照片共现情况、沟通发生的时间等），来判断某人是否属于家人或亲密朋友。由于系统的设计目标是把各种数据点串联起来，它甚至能仅凭元数据与时间戳推断出隐私信息，比如家庭中孩子的年龄、他们所在的学校或托儿所身份，以及设备使用者的规律性睡眠时间。

尽管这种监控的深度令人震惊，但几乎没有官方文档解释这些守护进程如何工作，或用户如何管理系统收集到的数据。删除机制也进一步加剧了隐私风险：当用户删除某个联系人后，实时图谱中的对应记录会被替换，但相关数据往往仍会留在 expired_stable_graph 表中，这意味着推断出的关系以及敏感推断可能会在用户以为已经移除之后很久仍然存在。

为帮助用户理解系统从自己身上推断出了什么，研究人员开发了一款名为 iKnow 的开源工具。该工具提供本地仪表板，允许用户直接在自己的设备上浏览个人知识图谱，让人得以看清操作系统在个人、群组与日常活动之间建立的连接。通过这种可视化，工具揭示了一个现实：Apple 的个人知识图谱本质上是把个人生活集中存放在一个高风险的仓库里，一旦访问控制出现问题就极易被利用，但对被画像的人而言，它在很大程度上仍然是不透明的。 Since macOS 13, Apple has quietly shipped a sophisticated background system consisting of two primary daemons whose collective purpose is to construct a detailed personal knowledge graph of the user. This system does not function in isolation, but instead aggregates data from a wide range of sources on the Mac, including Contacts, Calendar, Photos, Messages, Mail, location history, and Screen Time. By cross-referencing these inputs, the system creates a queryable, persistent database that tracks who you are, your family and social connections, your behavioral patterns, and even your physical movements and routines.

The heart of this mechanism is located within the directory ~/Library/IntelligencePlatform, which houses fifty-one SQLite databases. The most significant of these is graph.db, a triple store that organizes information using a vocabulary of 181 entity classes and 278 predicates. This graph records relationships between entities, assigning each connection a confidence score and documenting the source of the inference. While the system uses straightforward methods like copying data from contact cards or calendar entries, it also employs more advanced techniques, including CoreML-powered XGBoost classifiers, embedding models, and local LLMs to categorize relationships and identify individuals.

A particularly sensitive aspect of this system is its ability to infer kinship and social closeness without explicit user input. For instance, Apple includes pre-shipped XGBoost classifiers that analyze 144 distinct behavioral features, such as message frequency, photo co-occurrence, and communication timing, to determine whether someone is a family member or a close friend. Because the system is designed to link various data points, it can deduce private information, such as the age of a child in a household, the identity of their school or daycare, and the habitual sleep schedules of the device owner, simply by analyzing metadata and timestamps.

Despite the depth of this surveillance, there is almost no official documentation explaining how these daemons function or how users can manage the data they collect. Privacy is complicated further by the way the system handles deletions; when a user deletes a contact, the corresponding record in the live graph is replaced, but the data often lingers in an expired_stable_graph table, meaning that inferred relationships and sensitive inferences can persist long after the user believes they have been removed.

To help users understand what their Mac has inferred about them, researchers have developed an open-source tool called iKnow. This local dashboard allows users to browse their own knowledge graph directly on their machine, providing transparency into the connections the OS has drawn between individuals, groups, and daily activities. By providing this visibility, the tool highlights the reality that Apple's personal knowledge graph acts as a high-stakes, centralized store of personal life that is ripe for exploitation if access controls were ever compromised, yet it remains largely opaque to the person it is profiling.