Collaborative Research: Privacy-Sensitive Data Mining from Multi-Party Distributed Data

协作研究：来自多方分布式数据的隐私敏感数据挖掘越来越多的数据挖掘应用程序需要处理分布式的、可能是专有的、对隐私敏感的数据源。这导致了几种隐私保护数据挖掘技术的发展。其中许多算法使用随机化技术来扰乱数据并保护数据隐私，同时仍保证底层模式的不变性。这一建议首先指出，在许多情况下，流行的数据的朴素随机化(更具体地说，加法随机矩阵)可能会保留很少的数据隐私。提出了一个利用随机矩阵理论来理解这些数据掩蔽技术的框架，指出了现有隐私保护数据挖掘技术存在的问题以及解决这些问题的潜在研究方向，并建议开发一组在支持隐私敏感应用方面被证明是正确的随机化技术。它提出了一种基于随机投影的技术来从多方分布的数据中计算统计聚集。它还建议研究一种隐私敏感的分布式贝叶斯网络学习算法，用于类似的应用。建议的算法将使用PI在过去几年中一直在开发的一系列试验台来开发、测试和评估。

Collaborative Research: Privacy-Sensitive Data Mining from Multi-Party Distributed DataA growing number of data mining applications need to deal with data sources that are distributed, possibly proprietary, and sensitive to privacy. This has resulted in the development of several privacy-preserving data mining techniques. Many of these algorithms work using randomized techniques to perturb the data and preserve the data privacy while still guaranteeing the invariance of the underlying patterns. This proposal first points out that the popular naive randomization of the data (more specifically, additive random matrices) may preserve little data privacy in many cases. It proposes a framework to understand these data masking techniques using the theory of random matrices to shows the problems of some existing privacy-preserving data mining techniques and potential research directions for solving the problems.This proposal also suggests the development of a collection of randomized techniques that are provably correct in supporting privacy-sensitive applications. It proposes a randomized projection-based technique to compute statistical aggregates from multi-party distributed data. It also suggests research on a privacy-sensitive distributed Bayesian network learning algorithm for similar applications. The proposed algorithms will be developed, tested, and evaluated using a collection of testbeds that the PIs have been developing for the last several years.