This site uses cookies. More information presents the page The policy for cookies and similar technologies. CLOSE Close the warning message

Piotr Andruszkiewicz sustained his PhD thesis

On December 20, 2011: Piotr Andruszkiewicz sustained his PhD thesis: "Privacy Preserving Classification and Association Rules Mining over Centralised Data", supervised by Professor Marzena Kryszkiewicz.

This thesis is devoted to privacy preserving classification and association rules mining over centralised data distorted with randomisation-based methods which modify individual values at random to provide an expected level of privacy.
It is assumed that only distorted values and parameters of a distorting procedure are known during the process of building a classifier and mining association rules.
In this thesis, we have proposed the optimisation MMASK, which eliminates exponential complexity of estimating an original support of an itemset with respect to its cardinality, and, in consequence, makes the privacy preserving discovery of frequent itemsets and, by this, association rules feasible. It also enables each value of each attribute to have different distortion parameters. We showed experimentally that the proposed optimisation increased the accuracy of the results for high level of privacy.
We have also presented how to use the randomisation for both ordinal and integer attributes to modify their values according to the order of possible values of these attributes to both maintain their original domain and obtain similar distribution of values of an attribute after distortion.
In addition, we have proposed privacy preserving methods for classification based on Emerging Patterns. In particular, we have offered the eager ePPCwEP and lazy lPPCwEP classifiers as privacy preserving modifications of eager CAEP and lazy DeEPs classifiers, respectively.
We have applied meta-learning to privacy preserving classification. Not only have we used bagging and boosting, but also we have combined different probability distribution of values of attributes reconstruction algorithms and reconstruction types for a decision tree in order to achieve higher accuracy of classification. We have proved experimentally that meta-learning gives higher accuracy gain for privacy preserving classification than for undistorted data.
The solutions presented in this thesis were evaluated and compared to the existing ones. The proposed methods obtained better accuracy in privacy preserving association rules mining and classification. Moreover, they reduced time complexity of discovering association rules with preserved privacy.

Last modified: Monday, June 18, 2012 - 11:45:25 AM, Wacław Struk

x x News (7) - by publication date

‹‹ December 2011 ››
Mon Tue Wed Thu Fri Sat Sun
      1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 31