Utility Promises of Self-Organising Maps in Privacy Preserving Data Mining

Abstract

Data mining techniques are highly efficient in sifting through big data to extract hidden knowledge and assist evidence-based decisions. However, it poses severe threats to individuals’ privacy because it can be exploited to allow inferences to be made on sensitive data. Researchers have proposed several privacy-preserving data mining techniques to address this challenge. One unique method is by extending anonymisation privacy models in data mining processes to enhance privacy and utility. Several published works in this area have utilised clustering techniques to enforce anonymisation models on private data, which work by grouping the data into clusters using a quality measure and then generalise the data in each group separately to achieve an anonymisation threshold. Although they are highly efficient and practical, however guaranteeing adequate balance between data utility and privacy protection remains a challenge. In addition to this, existing approaches do not work well with high-dimensional data, since it is difficult to develop good groupings without incurring excessive information loss. Our work aims to overcome these challenges by proposing a hybrid approach, combining self organising maps with conventional privacy based clustering algorithms. The main contribution of this paper is to show that, dimensionality reduction techniques can improve the anonymisation process by incurring less information loss, thus producing a more desirable balance between privacy and utility properties.