Noisegate：为AI代理提供差分隐私保护的开源网关

Noisegate是一个开源差分隐私网关项目，使AI代理能够在数学保证下安全访问敏感数据，确保即使代理被操纵或恶意使用也无法泄露个人记录1。该项目通过三类经典隐私攻击演示验证了其保护效能：差分攻击、成员推断和重新识别均在系统中被击败，并在持续集成中进行回归测试1。项目的噪声机制经过验证，手动实现与行业标准库OpenDP在35项噪声尺度检查中精度匹配至1e-9，50万样本分布测试通过（D = 0.0014，p = 0.70）1。

在隐私预算管理上，Noisegate采用混合zCDP合成机制，相比传统的朴素预算会计提高了3倍效用：在相同隐私保证下，可支持308个查询而非100个1。使用默认参数设置（每个查询ε_q = 0.05，总预算ε = 5，δ = 10⁻⁶）时，对20行患者数据的平均查询误差为±12，对32,561行人口普查数据的平均误差仅为±60或0.2%1。系统为每个查询收取固定的per_query_epsilon，预算以用户身份、时期和数据集为键在本地持久化，重新连接不会刷新已消耗的预算1。

项目可作为MCP服务器与Claude Desktop集成，为AI代理提供结构化查询工具1。然而，开发者声明这是一个非生产级别系统，不应用于真实受管制数据（如受保护健康信息PHI、个人可识别信息PII），并依赖本地可信操作员来管理隐私参数1。

An open-source differential-privacy gateway called Noisegate has been introduced to enable AI agents to securely access sensitive data under mathematical guarantees1. The project ensures that even if agents are manipulated or used adversarially, individual records cannot be compromised1.

The system has been validated against three established privacy attacks: differencing, membership inference, and re-identification—all systematically defeated and subjected to regression testing in continuous integration1. The project's custom noise mechanism has been cross-verified against the industry-standard OpenDP implementation, matching noise scale calculations to 1e-9 precision across 35 checks and passing distribution tests on 500,000 samples with a Kolmogorov-Smirnov statistic of D = 0.0014 (p = 0.70)1. Using composite zCDP synthesis rather than naive budget accounting achieves a threefold utility improvement, supporting 308 queries under the same privacy guarantees compared to 100 queries with conventional methods1.

Under default parameters (per-query epsilon of 0.05, total budget epsilon of 5, and delta of 10⁻⁶), the system achieves mean errors of ±12 on 20-row patient datasets and ±60 or 0.2% on 32,561-row census datasets1. Each query consumes a fixed per-query epsilon allocation, with budgets persisted in a database indexed by identity, time period, and dataset; reconnections do not reset accumulated consumption1. The project integrates with Claude Desktop as an MCP server, providing AI agents with structured query tools1. The developers emphasize this is not production-ready and should not be deployed on regulated sensitive data such as protected health information or personally identifiable information, as it relies on local trusted operator oversight1.