Approval Sheet

Title of Dissertation: Multiplicative Data Perturbation for Privacy Preserving Data Mining Kun Liu, Doctor of Philosophy, 2007 Dissertation directed by: Dr. Hillol Kargupta Associate Professor Department of Computer Science and Electrical Engineering Recent interest in the collection and monitoring of data using data mining technology for the purpose of security and business-related applications has raised serious concerns about privacy issues. For example, mining health care data for the detection of disease outbreaks may require analyzing clinical records and pharmacy transaction data of many individuals over a certain area. However, releasing and gathering such diverse information belonging to different parties may violate privacy laws and eventually be a threat to civil liberties. Privacy preserving data mining strives to provide a solution to this dilemma. It aims to allow useful data patterns to be discovered without compromising privacy. In 2000, Agrawal and Srikant proposed the addition of i.i.d. white noise for privacy protection. However, Kargupta et al. pointed out that additive noise can be easily filtered off revealing a good approximation of the private data. This makes one wonder about the possibility of using multiplicative noise. This dissertation systematically investigated different multiplicative data perturbation techniques for privacy preserving data mining. These types of perturbation distort the private data by multiplying some random noise and only the perturbed version is released for data mining analysis. Extensive theoretical and experimental results were provided to support the following primary contributions. First, we examined the security issues of distance preserving data perturbation. This technique is potentially very useful in that some important data mining algorithms can be efficiently applied to the perturbed data and produce exactly the same results as if applied to the original data. However, the issue of how well the original data is hidden had not been carefully studied. We took a step in this direction by considering three types prior knowledge an attacker may have and use to design attack techniques to recover the original data. Our results offered insight into the vulnerabilities of distance preserving perturbation. Second, we explored a random projection-based data perturbation that preserves the inner products and Euclidean distances in the original data with high probabilities. We proposed a maximum a posteriori probability (MAP) estimate-based Bayes privacy model to quantify the privacy. Guidelines were offered for the data owner to control the privacy/accuracy tradeoff when perturbing the data. Theoretical analysis showed that this perturbation provides higher privacy protection than distance preserving perturbation, but with little loss of accuracy. MULTIPLICATIVE DATA PERTURBATION FOR PRIVACY PRESERVING DATA MINING