PRIORITIZED ANONYMIZATION FOR PUBLICATION OF SENSITIVE TRANSACTIONAL DATA

Abstract

Making better use of data is the main aim of data mining. Every organisation that posses data need not have data analysis or data mining expertise. So it necessitates sharing of data among organisation. But sharing the original data has many privacy concerns. Data-mining may help in revealing person specific information which may be sensitive. Privacy Preserving Data Publishing is a study of eliminating these privacy threats while providing most useful data. Privacy Preserving Data Publishing has received much attention from researchers in recent years. But most of the approaches proposed are for relational data with a fixed schema and low dimensionality. Also it has same number of sensitive attributes in every record, usually one. These existing techniques are not suitable for high dimensional transactional data as it leads to huge information loss. So separate approaches are required for them. There are only few works on sparse high dimensional data. In this dissertation, technique for Prioritized Anonymous Publication of Sensitive Transactional Data is proposed where different sensitive items can have different privacy requirement. This is an anonymized grouping method where general idea is to group transactions with close proximity and then associate each group to a set of sensitive values. For grouping we use nearest neighbour (NN) search using locality-sensitive hashing (LSH). We first convert our dataset to binary dataset, where d is the total number of items in the dataset, for efficient working of LSH. LSH will give reduced number of candidates for NN search. From this reduced candidate set required number of NN are selected which depends on privacy requirement of sensitive items present in the transaction. In proposed approach different sensitive items can have different privacy requirement. This is an extension of Anonymous Publication of Sensitive transactional Data where every sensitive item have same privacy requirement. Using same privacy for all sensitive items will reduce data utility as it demands same privacy for highly sensitive and less sensitive item. It will force us to compromise on privacy for better data utility. Hence we propose use of separate privacy requirement for different sensitive items which is a balanced approach. Experimental results show that this will improve data utility.