America.gov 的浏览器过滤器保留姓名和通用政府身份证件
被检查的客户端代码向 Rampart 的默认 keep-set 添加了四个类别，而模型卡则称通用政府标识符没有确定性的后备措施。
By Ryan Merket · Published · Updated
RUNTIMEWIRE INVESTIGATION — Original analysis
Original reporting by RuntimeWire, based on documents, testing.
为何重要
keep-set 决定检测之后会发生什么：对 America.gov 检查过的浏览器代码似乎将姓名、URL 和通用政府身份证明从 Rampart 的默认脱敏策略中排除。模型卡报告在对结构化政府身份证明的探测中召回率仅约为 67.6%，并表示这些标识符缺乏确定性的后备机制，而代码审查并未确定哪些会到达服务器或适用哪些服务器端控制措施。
Reporting record
Finding
America.gov's inspected browser code adds GIVEN_NAME, SURNAME, URL and GOVERNMENT_ID to Rampart's default keep-set, excluding spans with those labels from redaction or blocking in the inspected paths.
How we verified
Methods: documents, testing.
America.gov's client-side privacy wrapper, shared filter implementation and composer/PDF processing code; Rampart's published model card; and a synthetic-label check of four inspected span-selection functions.
Inspected America.gov's client-side JavaScript, traced the keep-set through typed-message privacy checks and PDF and feedback text-scrubbing paths, and checked four span-selection functions using synthetic labels. Compared the configuration with Rampart's published defaults. No personal identifiers were submitted.
Evidence
Reproduction
RuntimeWire independently reproduced the core finding.
Use synthetic labels to check the four inspected span-selection functions: GIVEN_NAME, SURNAME, URL and GOVERNMENT_ID were excluded from selection, while SSN, PASSPORT, DRIVERS_LICENSE, EMAIL, PHONE and BANK_ACCOUNT remained selected.
Rampart 模型卡的用途
模型卡为选择和部署模型的人提供一份简要记录，说明模型经过什么测试、在哪些方面不足以及其开发者认为适合的用途是什么。这有助于部署者判断模型是否适合特定工作以及在哪些方面需要额外测试或保障。模型卡并不为基于该模型构建的应用提供认证，也不记录已部署服务中的每一个决策。Hugging Face 的 model-card 指南 将预期用途、评估和限制列为核心内容；最初的模型卡论文 提出该格式以澄清适用场景并减少在模型表现较差的情境中的使用。
对于 Rampart，模型卡将一个实用的安全区分可视化：模型的检测界限与应用用以决定应当删除哪些内容的规则是分离的。Rampart 的模型卡指出，模型无法可靠地使用校验和验证许多政府身份证明，包括案件编号、类似 Medicare 的 ID、移民收据、护照和 licenses。在对结构化 ID 的测试中，模型的检测率约为 67.6%。该结果仅衡量模型本身，而非 America.gov 的完整系统。…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.