arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2103.08805cs.PLcs.CR
收藏

DDUO:差分隐私的通用动态分析

DDUO: General-Purpose Dynamic Analysis for Differential Privacy

Chike Abuah, Alex Silence, David Darais, Joe Near

摘要:

差分隐私允许对数据进行一般统计分析,并在个人层面提供隐私保护的正式保证。帮助数据分析师利用差分隐私的工具经常采用程序设计语言和库的形式。然而,设计用于差分隐私的组合验证的许多现有程序设计语言对程序员施加显著负担(以复杂类型注释的形式)。基于现有通用语言的隐私分析补充库支持更有用,但无法普及敏感性分析和隐私组合的端到端实施。我们介绍了DDUO,一个动态分析,以加强差分隐私。DDUO可供非专家使用:它的分析是自动的,不需要额外的类型注释。DDUO可以实现为现有的程序设计语言的库,我们提出了一个参考实现在Python中,具有适度的运行时开销在现实的工作负载。我们包括对现代机器学习程序中常用的几种数据类型、距离度量和操作的支持。我们还提供初始支持,用于跟踪流行Python库中数据转换的敏感性,以进行数据分析。我们形式化的DDUO系统的新的核心,并证明它的声音灵敏度分析通过一个逻辑关系度量保存。我们还通过各种案例研究来说明DDUO的可用性和灵活性,这些案例研究实现了最先进的机器学习算法。

英文摘要:

Differential privacy enables general statistical analysis of data with formal guarantees of privacy protection at the individual level. Tools that assist data analysts with utilizing differential privacy have frequently taken the form of programming languages and libraries. However, many existing programming languages designed for compositional verification of differential privacy impose significant burden on the programmer (in the form of complex type annotations). Supplementary library support for privacy analysis built on top of existing general-purpose languages has been more usable, but incapable of pervasive end-to-end enforcement of sensitivity analysis and privacy composition. We introduce DDUO, a dynamic analysis for enforcing differential privacy. DDUO is usable by non-experts: its analysis is automatic and it requires no additional type annotations. DDUO can be implemented as a library for existing programming languages; we present a reference implementation in Python which features moderate runtime overheads on realistic workloads. We include support for several data types, distance metrics and operations which are commonly used in modern machine learning programs. We also provide initial support for tracking the sensitivity of data transformations in popular Python libraries for data analysis. We formalize the novel core of the DDUO system and prove it sound for sensitivity analysis via a logical relation for metric preservation. We also illustrate DDUO's usability and flexibility through various case studies which implement state-of-the-art machine learning algorithms.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑