TWC: Medium: Privacy Preserving Computation in Big Data Clouds

隐私对于创造和创新的自由至关重要。有保障的隐私保护为行业创新、科学和工程发现以及新的生活体验和机会提供了前所未有的机会。在云中执行高效且保护隐私的大数据计算的能力为安全有效的数据分析提供了巨大潜力，例如使医疗保健应用程序能够使用个人的DNA序列提供个性化医疗，或使广告商能够通过挖掘用户的点击流和社交活动来创建有针对性的广告，而不会侵犯数据隐私。PrivacyGuard项目正在开发算法、系统和工具，在数据分析工作的整个生命周期内提供端到端的隐私保障。端到端隐私保证可以通过从经过处理的数据发布、执行的中间结果和分析作业的输出中了解一些原始敏感数据的难度来衡量。PrivacyGuard的最终目标是开发一个有条理的框架和一套技术，以确保分布式计算满足输入数据的预期隐私要求，并在执行期间和计算的最终输出中防止敏感模式的泄露。PrivacyGuard项目从三个方面提高了对保护隐私的分布式计算的认识和理解：(1)它设计了正式的机制来制定数据所有者对每次数据发布的端到端隐私要求，例如，通过将每次数据发布与定义良好的使用范围相关联，以限制可以对发布的数据进行操作的数据分析模型和算法集。(2)开发了一套具有双重目标的执行隐私保护：审计和执行分布式计算过程中的隐私合规性，防止基于数据流的隐私侵犯，并保护输入隐私的合规性。(3)针对与挖掘产出相关的信息泄露，设计了一种主动的输出隐私保护方法，例如，利用差分隐私模型最大化数据隐私保障的上界，最小化数据效用损失的下界。PrivacyGuard项目是确保分布式大数据计算中端到端隐私的实用和系统实施框架的第一个努力。此外，通过将PrivacyGuard的研究与佐治亚理工学院大数据系统和分析课程的课程开发相结合，有助于新一代数据科学家成为隐私合规倡导者的教育和培训。

Privacy is critical to freedom of creativity and innovation. Assured privacy protection offers unprecedented opportunities for industry innovation, science and engineering discovery, as well as new life enhancing experiences and opportunities. The ability to perform efficient and yet privacy preserving big data computations in the Cloud holds great potential for safe and effective data analytics, such as enabling health-care applications to provide personalized medical treatments using an individual's DNA sequence, or enabling advertisers to create targeted advertisements by mining a user's clickstream and social activities, without violation of data privacy. The PrivacyGuard project is developing algorithms, systems and tools that provide end-to-end privacy guarantees over the life cycle of a data analytic job. The end-to-end privacy guarantee can be measured by how difficult one can learn about some of the original sensitive data from the sanitized data releases, the intermediate results of execution and the output of an analytic job. The ultimate goal of PrivacyGuard is to develop a methodical framework and a suite of techniques for ensuring distributed computations to meet the desired privacy requirements of input data, as well as protecting against disclosure of sensitive patterns during execution and in the final output of the computation.The PrivacyGuard project advances the knowledge and understanding of privacy preserving distributed computation from three perspectives: (1) It designs formal mechanisms to formulate a data owner's end-to-end privacy requirement for each data release, for example, by associating each data release with a well-defined usage scope to confine the set of data analytics models and algorithms that can operate on the released data. (2) It develops a suite of execution privacy guards with dual objectives: to audit and enforce privacy compliances during distributed computation against data-flow based privacy violations and to guard the compliance of input privacy. (3) It devises a proactive approach to output privacy against information leakages associated with mining output, for example, by leveraging differential privacy model to maximize the upper bound for data privacy guarantee and minimize the lower bound for data utility losses. The PrivacyGuard project is the first effort towards a practical and systematic implementation framework for ensuring the end-to-end privacy in distributed big data computations. Furthermore, by integrating the PrivacyGuard research with the curriculum development on big data systems and analytics courses at Georgia Institute of Technology, it contributes to the education and training of new generation of data scientists to be the privacy compliance advocates.