Privacy preservation in data mining through noise addition

Due to advances in information processing technology and storage capacity, nowadays huge amount of data is being collected for various data analyses. Data mining techniques, such as classification, are often applied on these data to extract hidden information. During the whole process of data mining the data get exposed to several parties and such an exposure potentially leads to breaches of individual privacy. This thesis presents a comprehensive noise addition technique for protecting individual privacy in a data set used for classification, while maintaining the data quality. We add noise to all attributes, both numerical and categorical, and both to class and non-class, in such a way so that the original patterns are preserved in a perturbed data set. Our technique is also capable of incorporating previously proposed noise addition techniques that maintain the statistical parameters of the data set, including correlations among attributes. Thus the perturbed data set may be used not only for classification but also for statistical analysis. Our proposal has two main advantages. Firstly, as also suggested by our experimental results the perturbed data set maintains the same or very similar patterns as the original data set, as well as the correlations among attributes. While there are some noise addition techniques that maintain the statistical parameters of the data set, to the best of our knowledge this is the first comprehensive technique that preserves the patterns and thus removes the so called Data Mining Bias from the perturbed data set. Secondly, re-identification of the original records directly depends on the amount of noise added, and in general can be made arbitrarily hard, while still preserving the original patterns in the data set. The only exception to this is the case when an intruder knows enough about the record to learn the confidential class value by applying the classifier. However, this is always possible, even when the original record has not been used in the training data set. In other words, providing that enough noise is added, our technique makes the records from the training set as safe as any other previously unseen records of the same kind. In addition to the above contribution, this thesis also explores the suitability of pre-diction accuracy as a sole indicator of data quality, and proposes technique for clustering both categorical values and records containing such values.