BIGDATA: F: DKA: Scalable, Private Algorithms for Continual Data Analysis

出于同样的原因，大数据正在改变现代生活，它也对隐私和个人信息的控制构成了深刻的威胁。与大数据相关的一个主要挑战是，在不损害其中包含数据的个人隐私的情况下，实现对复杂数据集的统计分析。解决这一挑战既是必要的，因为由于隐私问题，对许多数据源的访问受到限制，也是困难的，正如对理应匿名的数据的大量攻击所表明的那样。这个项目将调查针对时变数据集的私有的、连续的分析的算法的设计和限制。也就是说，它将研究在收集数据集时(例如，以来自网络的数据流或长期社会学研究的形式)发布有关数据集的信息的算法。这项研究将推动对“大数据”--海量的、复杂的、时变的--数据的私下分析。如果成功，该项目将提供使能技术，促进在敏感数据访问受到保密问题限制的领域进行研究。该项目将专注于设计满足差异隐私的算法--差异隐私是计算机科学和相关领域广泛研究的一种严格的隐私概念。连续发布的隐私影响仍然知之甚少，在最近关于私人数据分析的大量文献中开发的算法中，相对较少的算法允许高精度的连续发布。该项目的两个主要推动力是(1)用于“持续发布”模型的算法，以及(2)用于提供更强隐私保障的“本地”模型的算法。这项工作将为这些模型提供新的算法设计技术和对算法复杂性理论限制的理解。这项研究将在学习理论、统计推理和流算法等相关领域取得进展。该项目还将包括旨在扩大研究影响的教育、外联和劳动力培训活动。有关更多信息，请参阅该项目的网站：http://www.cse.psu.edu/~asmith/projects/continual/

For the very same reasons that big data is transforming modern life, it also presents a profound threat to privacy and the control of personal information. A major challenge associated with big data is to enable statistical analysis of complex data sets, without compromising the privacy of the individuals whose data they contain. Addressing this challenge is both necessary, since access to many data sources is restricted due to privacy concerns, and difficult, as numerous attacks on supposedly anonymized data demonstrate. This project will investigate the design and limitations of algorithms for the private, continual analysis of time-varying data sets. That is, it will study algorithms that release information about a data set as it is collected (say, in the form of a data stream from the web, or a long-term sociological study). The research will advance the state of the art in the private analysis of "big" -- massive, complex, time-varying -- data. If successful, the project will provide enabling technologies that facilitate research in areas where access to sensitive data is limited by confidentiality concerns.The project will focus on the design of algorithms that satisfy differential privacy -- a rigorous notion of privacy that is widely studied in computer science and related fields. The privacy implications of sequential releases are still poorly understood, and relatively few of the algorithms developed in the extensive recent literature on private data analysis allow for sequential releases with high accuracy. The two major thrusts of the project are (1) algorithms for the "continual release" model, and (2) algorithms for the "local" model, which offers even stronger privacy guarantees. The work will provide novel algorithmic design techniques and understanding of complexity-theoretic limitations of algorithms for these models. The research will entail advances in related areas such as learning theory, statistical inference and streaming algorithms. The project will also include educational, outreach and work-force training activities designed to broaden the impact of the research.For further information see the project web site at: http://www.cse.psu.edu/~asmith/projects/continual/