面向最优效用的机器学习隐私模型

发布时间：2018-10-14 17:04

【摘要】：由于各种技术的发展,隐私问题成为当下很高的需求,这是因为越来越多的个人数据通过不同的组织,设备,和物联网(IOT)被分享,为个人数据的隐私与使用带来了风险,尤其是实现保密性的同时这样数据集的效用却在递减,因此,新的问题出现在要尽可能对一个更大的数据集保持数据隐私的同时保留多效用。在本研究中,建立了一个新的隐私模型,该模型使用Flash排序算法与k_匿名结合C4.5分类的机器学习技术,旨在保护隐私数据的同时,保持数据集的最优效用。该方法的第一步是在拥有30162条记录和属性的统计数据集上运用强大的数据隐私权授予技术,并使用闪光的排序算法选择最佳K-匿名数据集,设置隐私级别为2,然后通过C4.5分类过程使数据集达到尽可能多的效用。进一步,通过减少统计数据集的一半大小(共15081条记录),然后采用同样的方法进行测试。然后减少属性的数量,采用同样的算法进行测试。这项研究的结果揭示了一些重要的结果,与其他研究相比,本文的方法能够保持数据的准确性,结果显示数据集达到了90.77%的效用,此外,当数据集的大小减小为一半时,产生的数据效用的损失仅为0.5%,当数据集的属性减小时,效用损失仅为2.28%。而在数据集的大小较大时,与原来的非匿名数据集相比,损失为1.24%。虽然它提供了一种精度较高的结果,但未能在更大的数据集上取得最大的期望,研究结果表明,我们的方法在降低数据集的数量时,可以提供最低的效用值与属性约简数。该研究预测,改变隐私方法和使用不同类型的分类器在将来会产生更好的结果,尤其是在处理更大的数据集时。
[Abstract]:Because of the development of various technologies, privacy issues have become a high demand, because more and more personal data are shared through different organizations, devices, and the Internet of things, which brings risks to the privacy and use of personal data. Especially when confidentiality is realized, the utility of data set is decreasing. Therefore, the new problem is to keep the privacy of a larger data set while preserving multi-utility. In this study, a new privacy model is established. The model uses Flash sorting algorithm and K _ anonymity combined with C4.5 classification technology to protect privacy data while preserving the optimal utility of data sets. The first step of this method is to use powerful data privacy grant technology on statistical data sets with 30162 records and attributes, and select the best K-anonymous data set using flash sorting algorithm. Set the privacy level to 2, and then make the dataset as useful as possible through the C 4.5 classification process. Further, by reducing the size of the statistical data set by half (15081 records), then using the same method to test. Then reduce the number of attributes, using the same algorithm for testing. The results of this study reveal some important results. Compared with other studies, the method in this paper is able to maintain the accuracy of the data. The results show that the data set is 90.77% useful, and when the size of the data set is reduced to half, The loss of data utility is only 0.5. When the attribute of data set is reduced, the loss of utility is only 2.28. When the size of the data set is large, the loss is 1.24 when compared with the original non-anonymous data set. Although it provides a high precision result, it can not obtain the maximum expectation on a larger data set. The results show that our method can provide the lowest utility value and attribute reduction number when reducing the number of data sets. The study predicts that changing privacy methods and using different types of classifiers will produce better results in the future, especially when dealing with larger data sets.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP309;TP181

【相似文献】